Remote sensing-based detailed wetland classification: a review of advances from 2020 to 2025

Abstract

Wetlands play an irreplaceable role in maintaining ecological balance and conserving biodiversity. However, driven by both natural and human factors, vast wetland areas worldwide are being rapidly converted into agricultural or urban land, leading to a sharp decline in their extent and a degradation of their quality. Against this backdrop, remote sensing technology, with its unique advantages in macro-level monitoring and multi-temporal dynamic capture, provides indispensable technical support for the precise identification, classification, and change detection of wetlands, making it an essential tool for wetland monitoring, restoration assessment, conservation planning, and SDG-related evaluation. This review examines recent advances in remote sensing for detailed wetland classification over the past 5 years. It elaborates on commonly used remote sensing data sources, local and international wetland classification standards, diverse classification methods, accuracy evaluation metrics, and prospects. This review provides a comprehensive reference to remote sensing-based wetland classification studies and applications.

1 Introduction

As unique and vital ecosystems globally, wetlands feature diverse vegetation types and hold significant ecological value in maintaining ecological balance, regulating climate, and conserving biodiversity. Under the framework of the 2030 Agenda for Sustainable Development (Nations, 2015), wetlands are directly or indirectly linked to numerous Sustainable Development Goals (SDGs) indicators, including Clean Water and Sanitation (SDG 6), Life Below Water (SDG 14), Life on Land (SDG 15), and Climate Action (SDG 13) (Wetlands, 2018). However, due to the combined impacts of natural disturbances and human activities, global wetlands are experiencing area shrinkage, type conversion, and functional degradation (Mahdavi et al., 2018). Accurately understanding the composition, spatial distribution, and dynamic changes in wetland types is a prerequisite for assessing their ecological functions, formulating effective conservation strategies, and quantitatively evaluating the aforementioned SDG indicators (Hu et al., 2021).

Detailed wetland classification facilitates quantitative assessment of SDG-related indicators, thereby advancing ecological governance. Traditional wetland classification methods face limitations, including data acquisition challenges, labor-intensive ground surveys, low efficiency, and restricted spatial scales (Gallant, 2015). Remote sensing, however, offers advantages including macro-scale coverage, periodic observation, and multidimensional data acquisition. It enables rapid and accurate extraction of land-cover information in target areas, emerging as a key tool for overcoming these constraints (Zhang et al., 2013). However, wetlands are highly heterogeneous and strongly spatiotemporally dynamic land cover types shaped by hydrological processes, vegetation phenology, and human disturbance (Wen et al., 2025; Zhang X. et al., 2023; Luo et al., 2023). Detailed wetland classification generally refers to the remote sensing-based mapping of wetland subclasses below broad wetland categories, such as vegetation communities and hydrological subclasses, rather than simple wetland/non-wetland discrimination. Because these subclasses often exhibit subtle spectral, structural, and temporal differences, detailed wetland classification entails greater technical challenges while providing higher scientific value. Consequently, it has spurred extensive research worldwide, with significant methodological advances documented across diverse regions (Luo K. et al., 2025; Mahdianpari et al., 2018; Mahdianpari et al., 2020a; Mahdianpari et al., 2020b). Current approaches use object-oriented methods and multisource remote sensing data fusion, combined with prior knowledge (Igwe et al., 2023). The difficulty is that different wetland types may share similar spectral characteristics, while the same type can exhibit substantial variability under different seasons and hydrological conditions.

AI-driven wetland remote sensing classification has progressed from coarse-grained classification toward fine-grained, feature-level recognition (Wu et al., 2025), substantially improving identification capability and mapping accuracy in complex wetlands (Effah et al., 2025). However, researchers face several challenges, including high image acquisition costs, difficulties with image fusion, high annotation costs, and difficulties with feature extraction. Two constraints are particularly prominent. The first is limited training samples (Jamali et al., 2022a): wetlands typically occupy a small proportion of land cover, and detailed classifications are often spatially patchy, resulting in scarce training data and, consequently, poorer performance for wetland classes. The second issue is poor cross-regional generalization, which is especially pronounced in wetlands (Huang et al., 2023a). Mosaic patterns and transition zones (e.g., emergent/submerged vegetation and mudflats) increase intra-class variability while reducing inter-class separability, and even minor fluctuations in water level and moisture can shift the observed spectral signatures. Meanwhile, strong heterogeneity in hydrodynamics (O’Neil et al., 2020), tidal forcing (Yang X. et al., 2022), inundation frequency, and phenological rhythms induces substantial domain shifts (Xin et al., 2024), making models trained in one study area difficult to transfer to others (Tuia et al., 2016).

Detailed wetland classification remains a technically demanding task with distinct scientific challenges. Therefore, this paper aims to systematically review recent advances and synthesize a technical framework for fine-scale wetland classification using remote sensing. The paper first elucidates the remote sensing data sources and processing for wetland classification, then describes the technological evolution from traditional methods to deep learning models. It focuses on key aspects, including multisource data fusion, detailed wetland classification standards, and the advantages and limitations of classification models. The paper analyzes the primary challenges currently faced. It concludes with an outlook on future research directions, providing references for the precise management of wetland ecosystems, the assessment of SDG indicators, and ecological governance practices and conservation research.

2 State-of-the-art and technical framework for detailed classification of wetlands

To objectively and visually reveal recent research topics and evolutionary trends in fine-scale wetland classification using remote sensing, this study used WOS (Web of Science) databases to generate keyword co-occurrence and timeline analyses with CITESPACE software (Chen, 2006), covering global wetland remote sensing literature to contextualize China’s research within international trends. Detailed screening rules and parameter settings are described in Supplementary Appendix 1. The keyword co-occurrence map is shown in Figure 1. The figure indicates that current research topics primarily focus on the detailed mapping of wetland vegetation, the evolution of landscape patterns, ecological restoration assessments, and deep learning classification methods. At the data and methodology levels, the Google Earth Engine (GEE) platform, machine learning methods such as random forests, and deep learning frameworks have emerged as major technological drivers of recent advances in wetland classification. Temporal analysis indicates that since 2022, deep learning methods have gradually emerged as a research focus. By 2025, dynamic monitoring of coastal wetlands had emerged as a Frontier research topic. Improved data accessibility has lowered barriers to large-area processing and reproducible workflows, thereby accelerating the iterative advancement of methods such as random forests and deep learning (Gorelick et al., 2017; Belgiu and Drăguţ, 2016; Zhu et al., 2017). Additionally, feature extraction, accuracy, and temporal information mining are consistently prominent areas in this field.

FIGURE 1

The CiteSpace timeline of research topic evolution is illustrated below (Figure 2). It reveals that research on remote sensing classification of wetlands has undergone a multidimensional evolution over the past 5 years. Technologically, the focus has shifted from traditional machine learning algorithms (e.g., random forests) to deep learning models, represented by convolutional neural networks, and has further extended to refined feature extraction techniques, such as semantic segmentation. Platform support has leveraged cloud computing platforms, such as GEE, to advance automated data processing and scalable analysis. Research subjects encompass typical wetland systems, including salt marshes, coastal wetlands, and wetlands on Chongming Island, reflecting broad attention to diverse geographic settings. Application directions closely align with topics such as climate change responses, vegetation index construction, and the improvement of wetland classification accuracy. Collectively, these trends demonstrate an international research trajectory characterized by technological intelligence, cloud-based platforms, and refined research scenarios, within which Chinese studies have made substantial contributions, particularly in coastal wetland monitoring and alpine wetland mapping.

FIGURE 2

By synthesizing and reviewing the literature, this paper proposes an overall framework for detailed wetland classification (Figure 3). Multisource remote sensing data—optical, hyperspectral, SAR, and LiDAR—serve as inputs (Section 3). The workflow typically includes radiometric calibration, atmospheric correction, geometric correction and co-registration, data fusion, and study-area subsetting. From the preprocessed imagery, key features are derived, including spectral responses, texture and spatial metrics, temporal dynamics, and phenological indicators (Section 5.1). Classification models are then developed using these features, ranging from traditional machine-learning algorithms to deep-learning approaches (Sections 5.2, 5.3). Model performance is assessed with confusion-matrix-based metrics, including overall accuracy and the Kappa coefficient (Section 6). Once accuracy targets are met, detailed wetland maps are produced. The remainder of this paper is organized around the workflow.

FIGURE 3

3 Wetland remote sensing data sources and processing

3.1 Optical remote sensing images

Optical remote sensing passively receives reflected radiation across various bands, including visible, near-infrared, shortwave infrared, and thermal infrared, using optical sensors to acquire information about the Earth’s surface. It encompasses three types: panchromatic, multispectral, and hyperspectral imagery (Feng, 2023), with core distinctions lying in spectral dimensions and spatial resolution.

3.1.1 Panchromatic images

Panchromatic imagery consists of single-band grayscale images spanning the 500–750 nm wavelength range. While offering high spatial resolution, it cannot display color information (Feng, 2023). Therefore, in wetland classification, the panchromatic band is typically fused with multispectral data to generate color imagery. This approach combines the advantages of high resolution and multispectral coverage to enhance classification accuracy.

3.1.2 Multispectral remote sensing images

Multispectral remote sensing imagery features a limited number of broad spectral bands (typically 3–15 bands) that provide color information and diverse spectral combinations. Common satellite series include Landsat 8/9, Sentinel-2, the Gaofen series, WorldView-3, and other high-resolution data sources.

Landsat provides 11 bands: Bands 1–7 (visible, near-infrared, and shortwave infrared) and Bands 9–11 (thermal infrared) at 30-m resolution, while Band 8 (panchromatic band) offers 15-m resolution. Its strong continuity, extensive coverage, and 16-day revisit cycle make it suitable for long-term wetland monitoring at 30-m multispectral resolution, with the 15-m panchromatic band used for pan-sharpening (Yang M. et al., 2025). Landsat has been extensively used for the classification and dynamic monitoring of diverse wetland types, including coastal wetlands in the Yangtze River Delta, alpine wetlands on the Qinghai-Tibet Plateau, Ruoergai Wetland, subtropical coastal zones, and intertidal wetlands (Zheng et al., 2023; Peng et al., 2023; Chen et al., 2024; Xue et al., 2024).

Sentinel-2 offers 13 bands spanning the visible, near-infrared, and shortwave infrared spectral ranges (Tian et al., 2019), with a 5-day revisit cycle. Its resolutions include 10 m (visible and near-infrared bands), 20 m (red edge, near-infrared, and shortwave infrared bands), and 60 m (aerosol, water vapor, and cirrus bands). A key advantage of this data is the inclusion of red edge bands (B5, B6, B7, and B8A), which are sensitive to chlorophyll content, canopy structural condition, and vegetation-type discrimination (Sun et al., 2024). Red edge indices [e.g., NDRE (Huo and Niu, 2024), REMI (Chen et al., 2023)] not only enable sensitive monitoring of wetland vegetation growth but also aid in distinguishing wetland vegetation types (e.g., emergent versus submerged plants). Sentinel-2’s unique spectral information has demonstrated high accuracy in detailed classification studies of urban wetlands (Pan et al., 2024), coastal wetlands (Zong et al., 2021; Wang et al., 2024; Liu et al., 2021; Lin and Guo, 2024), the Yellow River Delta (Zhang et al., 2019), and other complex wetland environments (Lou et al., 2024).

The Gaofen (GF) satellite series provides core data support for detailed wetland research through its multisource data with high temporal, spatial, and spectral resolution (Qian H. et al., 2024; Sun, 2024; Zhou, 2024). Data types are primarily categorized into optical remote sensing (GF-1 to GF-7, with panchromatic resolution of 0.8 m and multispectral resolution of 3.2 m) and microwave remote sensing (GF-3). Among these, GF-2 leverages its sub-meter spatial resolution, combined with data fusion algorithms, to optimize classification, making it suitable for the detailed identification of small-scale features, such as small wetland patches, fragmented vegetation, and narrow channels (Gao et al., 2024). The GF-6 satellite carries a red edge band and a wide-swath imager. Its red edge band offers superior discrimination of wetland vegetation, filling gaps in traditional multispectral vegetation monitoring and proving suitable for large-scale wetland vegetation classification and boundary extraction (Huang et al., 2024).

3.1.3 UAV-based remote sensing

UAV (Unmanned Aerial Vehicle) represents a remote sensing platform rather than a specific image type. It can carry a variety of imaging sensors, including RGB, multispectral, hyperspectral, and thermal sensors, to provide flexible and task-specific observations for wetland mapping. UAV-Based remote sensing, provides crucial supplementary data usually conducted from low-altitude platforms, provides ultra-high spatial resolution data for detailed wetland classification (Zhu et al., 2022). With centimeter-level spatial resolution and high operational flexibility, UAV platforms overcome limitations such as long satellite revisit cycles and cloud cover, providing enhanced capabilities. They provide critical technical support for analyzing internal structures within specific wetland patches (e.g., vegetation cover, water depth distribution) and for achieving centimeter-level identification of key species (Tang et al., 2025; Huang et al., 2023b). This technical advantage has been widely validated in various wetland research settings (forest, coastal, and karst wetlands) (Zhang Y. et al., 2023; Bhatt and Maclean, 2023; Laporte-Fauret et al., 2020).

3.1.4 Hyperspectral remote sensing images

Hyperspectral remote sensing imagery captures detailed spectral characteristics through tens to hundreds of contiguous narrow bands (bandwidth <10 nm). Representative hyperspectral datasets include EOS AM-1 hyperspectral data, China’s Environmental Satellite HJ-1, Zhuhai-1 OHS-1, Ziyuan-1 02D, and GF-5 remote sensing data (Liu K. et al., 2022; Na et al., 2021). Due to their high spectral resolution, hyperspectral remote sensing images can effectively distinguish between land cover types with similar spectral characteristics (Feng, 2023; Han et al., 2023; Liu Y. et al., 2023; Long et al., 2025; Wang et al., 2023b; Zhang M. et al., 2025; Zhu et al., 2025), making it suitable for scenarios requiring precise differentiation of wetland vegetation. However, it involves large data volumes, high processing costs, and susceptibility to atmospheric effects, typically necessitating radiometric calibration. Classification accuracy and reliability can be enhanced through dimensionality reduction and feature-selection methods, such as PCA, manifold learning, and RF-based feature importance ranking, or by integrating radar imagery to improve all-weather monitoring capabilities (Sun et al., 2021).

3.2 Active remote sensing data

3.2.1 Synthetic aperture radar (SAR)

SAR is sensitive to surface roughness, moisture, and inundation, with key parameters for wetland classification including polarization (VV for water surface, VH for vegetation volume scattering; quad-polarization from GF-3 provides comprehensive scattering information) (Wang et al., 2023b; Yao et al., 2022), wavelength (C-band Sentinel-1 for herbaceous wetlands, L-band ALOS-2 for forested wetland penetration) (Zhao et al., 2022), and incidence angle (steeper angles enhance inundation detection beneath canopies). Speckle filtering (Lee, Frost) preserves edges while reducing noise (Yao et al., 2022). Tidal-stage effects significantly alter backscatter in coastal wetlands and must be considered in multi-temporal fusion (Zhao et al., 2022; Li et al., 2025a). Sentinel-1 (6-day revisit, VV/VH) is widely used (Wang et al., 2023b; Zhao et al., 2022; Li et al., 2025a).

3.2.2 Light detection and ranging (LiDAR)

LiDAR provides three-dimensional terrain data for wetland microtopography (sub-meter elevation variations controlling hydrological connectivity) (Martins et al., 2020), elevation gradients (discriminating high marsh from low marsh), and canopy height models for mangrove species discrimination (Martins et al., 2020). Cost and coverage limitations favor fusion with optical or SAR data (Guo et al., 2023; Martins et al., 2020).

3.3 Multisource remote sensing data fusion

Multisource remote sensing image fusion integrates information acquired by different sensors with complementary spatial resolution, spectral sensitivity, imaging mechanisms, and revisit cycles to produce fused data that contains more detailed information than any single source (Zhang, 2010). According to the fusion level, the process can be classified into three types: pixel-level, feature-level, and decision-level fusion. Pixel-level fusion emphasizes information enhancement. Feature-level fusion focuses on jointly representing features extracted from multiple data sources and learning from their collaboration. Decision-level fusion improves robustness by integrating multiple algorithms into a final decision (Zhang, 2010).

In wetland classification, optical imagery is generally more sensitive to the spectral differences between water bodies and vegetation. In contrast, SAR is more responsive to surface roughness, moisture content, and inundation conditions. Time-series observations can further characterize seasonal hydrological dynamics and vegetation phenological fluctuations. Consequently, image fusion has become the mainstream approach for enhancing the accuracy of wetland classification (Ren et al., 2020). It enables the construction of a more discriminative feature space (Wu et al., 2024). Specifically, the fusion of optical and radar imagery (e.g., Sentinel-1/2 time-series data) (Wu et al., 2022; Sahour et al., 2021) and the integration of hyperspectral and radar data (Tu et al., 2021; Guo et al., 2023) have significantly expanded the dimensionality of land cover features (Luo B. et al., 2025). Building upon this foundation, the further incorporation of spectral characteristics, spatial contextual information, and phenological patterns enables deep feature-level synergy. This approach not only improves classification accuracy (Huang et al., 2023c) but also enhances the ability to capture wetland dynamics (Ming et al., 2023). It should be noted that current multisource data fusion still largely relies on “simple superposition” and lacks deep integration driven by cross-modal feature collaboration (Yue et al., 2025). Moreover, using more inputs is not necessarily better; effectiveness depends on whether complementary cross-modal features can be explicitly captured (Gao et al., 2021).

3.4 Google earth engine (GEE): a cloud computing platform

The emergence of the GEE cloud computing and data-processing platform in recent years has significantly advanced wetland remote sensing research (Gorelick et al., 2017) by providing access to multisource time-series imagery and scalable computing resources for compositing, feature extraction, and classification. GEE has driven the field away from relying on local computing and fragmented data toward platform-based, integrated, and reproducible cloud-based research workflows (Tamiminia et al., 2020; Amani et al., 2020). It integrates vast amounts of free, multisource time-series imagery (e.g., Landsat, Sentinel-2, MODIS) with global-scale cloud computing capabilities, effectively resolving long-term bottlenecks in data acquisition, storage, and computational efficiency in wetland research (Wu et al., 2024; Demarquet et al., 2024; Jiang et al., 2023). On the one hand, the efficient processing of long time-series data on GEE has enabled large-scale, high-frequency wetland mapping and dynamic monitoring at urban, regional, national, and even global scales (Xue et al., 2024; Tootchi et al., 2019). On the other hand, its script-based cloud computing environment has further facilitated the deep integration of advanced algorithms (e.g., machine learning models such as random forests) into wetland classification, improving methodological reproducibility and result reliability, thereby accelerating the credibility of iterative model development (Pan et al., 2024; Wang et al., 2024; Sahour et al., 2021; Turnbull et al., 2024; Yang Z. et al., 2022; Cheng, 2022; Fu et al., 2022).

3.5 Comparison of remote sensing data sources

In wetland remote sensing classification studies, the selection of remote sensing data sources has a critical influence on classification accuracy and application effectiveness. Table 1 summarizes the data characteristics, advantages, limitations, and application scenarios of commonly used remote sensing imagery in wetland monitoring, providing a reference for subsequent research to optimize remote sensing classification methods and strategies. Overall, multispectral data such as Landsat and Sentinel-2 are suitable for monitoring large-scale wetland landscape patterns and long-term temporal dynamics, particularly at the macro-scale. For studies of small wetlands or detailed mapping, high-resolution imagery such as Gaofen-2 (GF-2) and WorldView is recommended. In heavily cloud-covered and rainy regions, radar data from sources such as Sentinel-1 and Gaofen-3 (GF-3) become indispensable remote sensing tools due to their strong penetration capabilities. For studies focusing on physiological and biochemical parameters of wetland vegetation or water quality indicators, hyperspectral data, such as those from Gaofen-5 (GF-5), should be explored. Multisource data fusion represents both a research hotspot and a future trend. By leveraging the strengths of optical and radar data, the accuracy of wetland classification and monitoring capabilities can be significantly enhanced.

4 Wetland classification criteria

After acquiring study-area imagery, subsequent processing and classification must adhere to specific wetland classification standards. The adopted classification framework defines the mapping targets and thematic detail of wetland classes and therefore influences subsequent choices of remote sensing data, feature extraction strategies, classification methods, and accuracy assessment approaches. Multiple classification systems are used internationally. The Ramsar Convention established an international framework for a wetland classification system (Ramsar, 2024). The US Fish and Wildlife Service (USFWS) system categorizes wetlands into five levels to support consistent mapping and interpretation (Service, 2023). In global or cross-regional wetland classification, the Ramsar framework is frequently used as a typological reference, with studies often operationalizing it by using remotely observable major wetland types to enhance cross-regional consistency and class separability (Yang Z. et al., 2025; Yin et al., 2026; Zhang Z. et al., 2025; Zhang X. et al., 2024).

At the national or regional level, adaptations are common. For instance, China has established a classification system that categorizes wetlands into five major types: marsh wetlands, lake wetlands, river wetlands, coastal wetlands, and constructed wetlands, further subdivided into 28 distinct types (Administration, 2009). In contrast, the Ramsar framework and USFWS system emphasize hydrological regimes, vegetation assemblages and ecological mapping. Compared with international frameworks, the Chinese scheme is more inventory- and management-oriented and better suited to national applications, while the representation of hydrological dynamics is relatively simplified (Carlson et al., 2024). Owing to these different organizing principles, cross-system comparisons at the detailed level often lack strict one-to-one correspondence. For instance, China’s “herbaceous marsh” may correspond to multiple hydrological subclasses in the USFWS system. In contrast, the “intertidal forested wetland” in the Ramsar system is classified under the mangrove subclass of “coastal wetlands” in the Chinese scheme. Based on these differences, the choice of classification standard at the national level should be determined by the specific task and spatial scale (Mao et al., 2020; Zhou et al., 2024; Zhang J. et al., 2024).

As one of the most representative regions with comprehensive wetland coverage, China has seen extensive wetland studies (Figure 4) that address the major wetland types defined by different classification systems. As shown in the figure, the spatial distribution exhibits a highly diverse and complex pattern, ranging from high-elevation alpine marsh wetlands on the Tibetan Plateau and southwestern mountains, to coastal wetlands along the eastern coasts, to lake wetlands and urban wetlands around big cities, spanning multiple climate zones, ecological settings, and gradients of human disturbance. Consequently, detailed wetland classification in China faces more pronounced challenges arising from regional heterogeneity.

FIGURE 4

Existing research in China has primarily focused on the detailed subclasses summarized in Table 2. Although classification criteria are often adapted to local conditions, recurrent confusion patterns can be identified across regions and wetland types. Table 2, therefore, compiles representative Chinese study areas, their detailed subclass schemes, and the commonly reported confusion-prone subclasses and typical errors, providing a basis for a more unified framework and future cross-regional comparisons.

The primary challenges in China’s wetland detailed classification center on: i) mosaicked or mixed vegetation and blurred class boundaries, ii) weak separability at water–tidal-flat interfaces, and iii) spectrally similar herbaceous communities and human-modified land covers. To mitigate this confusion and misclassification, prior studies commonly introduce multisource features (e.g., dual-polarization SAR backscatter), incorporate object-based strategies, use ensemble learning or deep network architectures, and leverage sample transfer/domain adaptation to enhance inter-class discrimination; these methods are discussed in the next section.

5 Detailed wetland classification methods using remote sensing

Researchers have developed and explored a wide range of methods to enable more detailed wetland classification. Existing approaches can be grouped according to several commonly used criteria. Regarding whether labeled samples are required, methods are generally classified as supervised or unsupervised. In terms of the smallest mapping unit, methods can be categorized as pixel-based, object-based, or hybrid approaches that incorporate pixel decomposition. From the perspective of feature representation and learning paradigms, the methods can be broadly classified into manual feature methods, machine learning, and deep learning approaches (Zhang et al., 2018). Accordingly, this paper reviews and compares representative methods from the standpoint of feature representation and learning strategy.

5.1 Feature engineering and dimensionality reduction

Manual feature description methods rely on experts with extensive domain knowledge and practical experience to manually design and extract image features for wetland identification. This corresponds to the feature extraction and selection step in the technical workflow (Step 2 in Figure 3), including spectral features (e.g., NDVI, NDWI, and other vegetation/water indices that highlight spectral differences between wetlands and non-wetlands through band combinations); spatial relationship features (e.g., texture features based on grayscale co-occurrence matrices, object structural features after multiscale segmentation capturing spatial morphological differences in wetland landscapes), temporal sequence features (e.g., phenological fluctuations of wetland vegetation and seasonal patterns of water bodies in multi-temporal imagery), and auxiliary discriminative features (e.g., supplementary information such as topographic factors and phenological cycles) (Zhang et al., 2018).

With increasing resolution and dimensionality of remote sensing imagery, individual features struggle to represent the complex attributes of wetlands comprehensively. Multi-feature fusion has thus emerged as a key strategy, integrating spectral, spatial, temporal, and auxiliary multidimensional features to overcome the limitations of individual features. However, classification performance varies significantly across different feature combinations. Therefore, selecting effective feature combinations and constructing targeted feature sets remain critical steps in wetland classification.

As remote sensing data become increasingly high-dimensional, particularly hyperspectral imagery, dimensionality reduction techniques are often employed to reduce redundancy and computational complexity while preserving the most informative features. Among these, Principal Component Analysis (PCA) is one of the most widely used approaches in wetland remote sensing. PCA projects high-dimensional data onto a low-dimensional space via a linear transformation to capture primary variation while eliminating redundancy and noise.

In hyperspectral wetland imagery, PCA can compress hundreds of spectral bands into a small number of principal components, thereby enhancing the separability of wetland categories and reducing misclassification. Previous studies have demonstrated its effectiveness in reducing dimensionality for high-dimensional datasets and in improving classification performance through PCA-based feature selection (PAN and LIN, 2022; Martins et al., 2020). However, because PCA is a linear transformation method, it is limited in modeling nonlinear spectral mixing processes in highly heterogeneous wetlands. Consequently, nonlinear alternatives such as kernel PCA and manifold learning may yield further improvements in classification performance (Jarocińska et al., 2024; Xu S. et al., 2023). Nevertheless, PCA remains an efficient and widely adopted preprocessing technique for preparing inputs for deep learning-based wetland classification models (Liu H. et al., 2025).

5.2 Traditional machine learning classification

In traditional machine learning, shallow models such as Support Vector Machines (SVMs) and Random Forests (RFs) have been widely applied to wetland classification, with approaches categorized as pixel-based or object-oriented. Pixel-based methods often produce “salt-and-pepper” patterns in classification results and are particularly prone to misclassifying land types with similar spectral signatures. In contrast, object-based classification integrates geometric, textural, and spectral information, effectively mitigating these issues (Zhang et al., 2018) and generally achieving better performance than pixel-based methods (Fu et al., 2022; Du et al., 2021). The key step in object-based classification is multiscale segmentation, which captures features at different scales by repeatedly segmenting and merging scales, thereby improving the accuracy and robustness of the classification results. Because segmentation quality may be affected by land-cover confusion and suboptimal boundaries, it is crucial to set parameters appropriately and select suitable algorithms to obtain the best segmentation outcome. The following section introduces the main algorithms from three aspects.

5.2.1 Support vector machines

The core concept of Support Vector Machines (SVMs) is to identify an optimal separating hyperplane that maximizes the margin between different classes (Figure 5). SVM applications for detailed wetland classification—including deep feature extraction (Wang Z. et al., 2025) and combining maximum likelihood with Mahalanobis distance (Tang et al., 2025)—demonstrate advantages in classifying wetlands in the Yellow River Delta and the Yancheng coastal wetlands of Jiangsu Province. SVM can effectively handle high-dimensional features and small-sample settings and is often efficient to train (Chen et al., 2024); however, its feature representation capacity is limited, making it difficult to capture subtle textural differences along complex wetland boundaries (Zhang L. et al., 2025).

FIGURE 5

5.2.2 Decision trees

Decision trees represent decision rules and classification results through tree-like data structures (Figure 6). In wetland classification, leaf nodes represent a classification, while non-leaf nodes correspond to divisions based on specific attributes. The core challenge in constructing decision trees lies in selecting appropriate attributes for splitting the sample at each step. Previous studies (Liu Y. et al., 2023; Cheng, 2022; Zhuang and Kuang, 2024; Leishi, 2021) have combined optical imagery with object-oriented decision-tree models to achieve rapid, high-precision wetland identification and classification, demonstrating the model’s applicability across various wetland types, including riverine, lacustrine, and intertidal wetlands. A single decision tree is sensitive to noise and sample perturbations and can easily overfit; therefore, pruning or ensembling with other frameworks (e.g., two-stage schemes combined with random forests) is often used to improve generalization and robustness (Wang et al., 2023a).

FIGURE 6

5.2.3 Random forest

Random Forest (RF) is a classification algorithm based on ensemble learning, whose core lies in collective decision-making through the voting of multiple decision trees (Figure 7). RF can maintain strong adaptability and robustness in complex settings—characterized by high dimensionality, high inter-class spectral similarity, and noise interference—under limited hyperparameter tuning, low labeling cost, and small-sample conditions (Adam et al., 2010; Wang et al., 2019). Leveraging these advantages, RF is widely applied to large-area wetland classification (Zhang et al., 2018; Deng et al., 2023; Liu Y. et al., 2025; Shi et al., 2025) and subsequent vegetation assessment (Turnbull et al., 2024).

FIGURE 7

Numerous optimizations have been developed for RF, significantly enhancing model performance and detailed recognition capability. At the structural level, studies have employed cascaded RF classifiers for hierarchical feature learning, yielding high-accuracy results in wetland classification with the Multi-Granularity Cascade Forest (gf-Forest) and Patch-Based Cascade Forest (PBC) (Judah and Hu, 2022; Shu et al., 2024). Among these, the multiscale granularity cascaded forest (OGCF) is particularly representative and demonstrates notable advantages in both accuracy and computational efficiency in resource-constrained, multiscale, complex wetland classification scenarios (Liu H. et al., 2022). At the feature- and rule-fusion levels, combining random forests with recursive feature elimination algorithms optimizes features to enhance classification accuracy (Zhang et al., 2018; Shi et al., 2025). Coupling random forests with object-oriented knowledge rule decision models achieves fine-grained classification (Deng et al., 2023), while integrating color features into random forests improves recognition performance (Liu Y. et al., 2025).

5.3 Deep learning methods

Deep learning, a branch of machine learning, automatically extracts high-level features from raw data using multi-layer neural network architectures without human intervention. Compared to traditional shallow models (e.g., SVM, RF, decision trees), deep learning algorithms require larger datasets, support more complex models, demand higher computational resources during training, and involve longer training times, primarily relying on large-scale computing resources. In recent years, deep learning methods have achieved promising results in wetland classification (Yang Z. et al., 2022). Representative approaches include Convolutional Neural Networks (CNNs) for image recognition, Recurrent Neural Networks (RNNs) for processing sequential data, Generative Adversarial Networks (GANs) for handling sequences of varying lengths, and popular models such as Transformers and U-Nets.

5.3.1 Convolutional neural networks

Convolutional Neural Networks (CNN) are deep learning architectures designed for two-dimensional data such as raster imagery. They typically consist of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer, as shown in Figure 8. CNNs extract local spatial features, such as texture and shape, through convolutional operations, then reduce feature dimensions via pooling layers, with the final classification task performed by fully connected layers.

FIGURE 8

CNN has been widely applied in wetland classification and has achieved good performance (Wang et al., 2019; Jamali et al., 2021a), demonstrating strong adaptability in heterogeneous classification scenarios. In coastal wetlands, architectural modifications, attention mechanisms, and deep feature extraction can enhance discrimination of mixed vegetation and blurred boundaries (Cui et al., 2023; Jamali et al., 2021a; Zhang et al., 2023; Li et al., 2024a). In emergent vegetation zones, multi-temporal inputs and temporal feature optimization help mitigate spectral fluctuations caused by freeze–thaw cycles and phenological changes, thereby improving classification stability (Meng, 2019; Li K. et al., 2024). In urban wetlands, CNN models outperform traditional methods in identifying highly fragmented small patches and complex backgrounds (Jamali et al., 2021a; Yang R. et al., 2022). In addition, studies comparing the adaptability of different backbone networks (Yang R. et al., 2022) and employing model fusion for improved stability (Meng, 2019) further confirm the applicability of CNNs. CNN models exhibit robust capabilities in spatial feature extraction and representation and outperform traditional machine learning methods in classification accuracy and boundary delineation (Cui et al., 2023).

Deep Convolutional Neural Networks (DCNNs) represent an extension and structural optimization of classical CNNs, enhancing semantic feature learning through deeper hierarchical structures. A key branch of DCNNs is the Fully Convolutional Network (FCN), which eliminates connected layers to enable end-to-end pixel-level classification, making it highly suitable for detailed wetland classification. Previous studies have combined object-oriented techniques with FCN (Ma et al., 2022) and integrated multisource imagery to build DCCNN models (Jamali et al., 2021b), thereby significantly improving overall classification accuracy, particularly for subcategory identification in marsh wetlands.

CNNs consume substantial computational resources and demand high-quality labeled samples. As network depth increases, models are more prone to vanishing gradients and overfitting, which can limit performance in processing high-resolution remote sensing imagery. Transfer learning and semi-supervised learning can reduce reliance on large-scale manual annotation; alternatively, the models introduced below can improve the robustness and generalization of wetland classification through multiscale fusion and attention mechanisms.

5.3.2 Recurrent neural networks and variants

Recurrent Neural Networks (RNNs) excel at processing sequential data. By incorporating recurrent structures, each neuron receives input from both preceding layers and the current layer, making them particularly suited for sequence prediction and language modeling tasks, as shown in Figure 9. In wetland research, RNNs and their variants (e.g., LSTM, Bi-LSTM, TSTLN) can accurately capture the dynamic evolution of wetland ecosystems using remote sensing imagery or index time series data across different years and seasons. They are widely applied in mangrove extraction (Xue and Qian, 2023) and complex wetland classification (Radman et al., 2024), achieving significantly higher classification accuracy than traditional RNN models. To optimize wetland time-series analysis, RNNs are often combined with CNNs to leverage the latter’s strengths in extracting local textures and deep semantic features from images, thereby compensating for the limitations of either model alone (Radman et al., 2024; Xu M. et al., 2023).

FIGURE 9

Furthermore, integrating RNNs with Generative Adversarial Networks (GANs) effectively mitigates the scarcity of labeled samples (Chen et al., 2022), thereby enhancing classification accuracy and robustness in small-sample wetland scenarios (Jafarzadeh et al., 2022). When processing ultra-long sequence data, RNNs are constrained by their recurrent structure, prone to vanishing and exploding gradients, and not easily parallelizable. This results in low computational efficiency and prolonged training times.

5.3.3 Generative adversarial networks

Generative Adversarial Networks (GANs) pioneered a new paradigm for data-driven feature learning through an adversarial game between a generator and a discriminator (Figure 10). The generator aims to create synthetic samples that closely resemble the distribution of real data, while the discriminator assesses the authenticity of the input data. This mutual competition continuously enhances the model’s generalization capability, as shown in Figure 10.

FIGURE 10

GANs and their variants have been introduced into fine-grained wetland classification, mainly to address challenges such as sample scarcity, inter-class imbalance, and model performance optimization. For data augmentation, combining 3D-GAN with V-Transformer can maintain high accuracy even with reduced training samples (Jamali et al., 2022a), while expanding Suaeda salsa samples using conditional GANs can effectively improve classification accuracy in coastal wetlands (Xiong et al., 2022). For architectural design, researchers have developed semi-supervised GANs for small-sample classification (Liu, 2023). For framework construction, GAN-integrated deep learning frameworks have also emerged; for instance, WetNet achieved superior performance in large-scale wetland classification through spatiotemporal integration (Hosseiny et al., 2022). Overall, GANs show considerable potential for wetland classification, but because their training is sensitive to network structure and parameter settings, careful design and tuning remain necessary in practical applications.

5.3.4 Transformer models

Leveraging self-attention and encoder-decoder architectures, Transformers (Figure 11) enable efficient modeling of long-sequence wetland data and show clear advantages in hyperspectral and multimodal fusion (Sun, 2023; Qian S. et al., 2024). Addressing GPU memory bottlenecks in high-resolution remote sensing data processing, patch-based strategies have become standard preprocessing steps. This approach not only leverages the advantages of GPU-based parallel computing by cropping large images into smaller patches but also enhances the representation of fragmented wetland features and complex boundaries (Marjani et al., 2024; Lijia, 2025). Facing the common challenge of scarce labeled samples in coastal wetland classification, transfer learning is regarded as an effective and practical strategy. It reduces the demand for training data in the study area by enabling knowledge transfer (Xiong et al., 2022; Hosseiny et al., 2022). Transformer-enhanced architectures (Li W. et al., 2024; Zou et al., 2025; Liu W. et al., 2023) have demonstrated significant advantages in hyperspectral wetland classification. The spectral-spatial joint attention mechanism can characterize inter-band relationships and spatial neighborhood information, and effectively address spectral redundancy in hyperspectral data in complex areas such as tidal creek boundaries (Xin et al., 2023; Guo et al., 2022). At the same time, a hybrid CNN–Transformer can effectively integrate local texture features and global dependencies, thereby improving the recognition of complex classes such as mudflats and mixed vegetation in coastal wetlands (Li et al., 2025b).

FIGURE 11

5.3.5 U-net model

The U-Net architecture employs a symmetric encoder-decoder structure with skip connections (Figure 12). While the encoder downsamples to capture high-level abstract features, the decoder upscales to restore high-resolution spatial details. This unique “U-shaped” information flow enables the model to integrate multiscale feature information while preserving original image details. U-Net demonstrates strong adaptability in complex wetland landscapes (Huang et al., 2023b). Its skip-connection mechanism fuses shallow-level localization features with deep-level semantic features, significantly improving the accuracy for challenging wetland boundary segmentation tasks, such as water-vegetation transition zones. Data augmentation strategies effectively mitigate the scarcity of annotated wetland samples (Ronneberger et al., 2015).

FIGURE 12

Recent studies suggest that U-Net is being extended in two main directions: architectural refinement for high-resolution wetland mapping, and integration with attention/Transformer modules to strengthen global context modeling. Improved U-Net variants on Gaofen-2 imagery have shown advantages over classic backbones such as FCN and SegNet (Ma et al., 2023). For heterogeneous and temporally dynamic wetlands, Transformer-coupled designs using Sentinel multi-temporal inputs can better capture long-range dependencies, reporting higher overall accuracy and Kappa than Swin-UNet and DeepLabV3+, while also improving inference efficiency in some cases (Wang B. et al., 2025; Gou et al., 2025). In addition, attention-augmented U-Net models enhance the extraction of key wetland features and further improve classification performance (Li et al., 2024b). Overall, Transformer integration is driving U-Net toward a local–global dual-stream interpretation (Wang B. et al., 2025; Li et al., 2024b).

6 Accuracy evaluation methods and metrics

6.1 Evaluation methods

6.1.1 Accuracy evaluation method

The accuracy evaluation method quantifies model reliability by comparing classification results with independent, higher-quality reference data. A rigorous assessment requires explicit decisions on sampling design, response design, and analysis, with probability-based sampling and transparent protocols recommended to support statistically defensible inference (Congalton, 1991).

Validation samples should be selected using probability-based sampling schemes where possible, such as simple random or stratified random sampling, and should remain independent from training samples to avoid optimistic accuracy estimates. Stratified random sampling is particularly useful because detailed wetland subclasses are usually unevenly distributed, and rare classes may otherwise be underrepresented. In addition, validation samples should be spatially well distributed to reduce the influence of spatial autocorrelation. For large-area wetland mapping, area-adjusted accuracy assessment and confidence intervals or other uncertainty estimates can further improve the reliability and interpretability of the evaluation.

6.1.2 Confusion matrix evaluation method

The confusion matrix is a common and effective method for validating wetland classification. Rows represent reference classes, columns represent predicted classes, and the diagonal elements indicate the number of correctly classified samples. It provides an intuitive display of overall accuracy and levels of confusion (similarity/disparity), enabling analysis of primary causes of confusion and model refinement. For large-area products, it reports an estimated population error matrix and uncertainty (Foody, 2002).

6.2 Metrics

6.2.1 Kappa coefficient

The Kappa coefficient measures the agreement between the classification results and a random classification, reflecting the classifier’s performance. Generally, a Kappa coefficient greater than 0.8 indicates excellent classification results, suitable for delineating core areas within protected zones. Values between 0.4 and 0.8 indicate relatively reliable classification results, supporting the analysis of large-scale wetland dynamic change trajectories. Below 0.4, classification results are poor and unreliable. Compared to accuracy, the Kappa coefficient eliminates random interference and is suitable for evaluating wetland classifications with uneven sample distribution (Sun et al., 2020). However, the Kappa coefficient is highly sensitive to class prevalence and sample balance, particularly when dominant wetland classes occupy most samples. Therefore, it should be interpreted alongside class-specific metrics such as Producer’s Accuracy (PA), User’s Accuracy (UA), and F1 score when evaluating wetland classification results (Farhadpour et al., 2024).

6.2.2 Overall classification accuracy

Overall Accuracy (OA) is one of the most intuitive metrics for evaluating wetland classification precision, reflecting the proportion of correctly classified validation samples, or an area-adjusted estimate derived from validation samples. Generally, classification products with high OA are suitable for wetland monitoring; in practice, targeted manual interpretation and, where feasible, field verification of key patches are often required, whereas products with insufficient OA typically require model retraining or workflow redesign. OA provides a rapid assessment of classification performance, serving as a reference for subsequent optimization of classification processes and methods to enhance accuracy. However, OA is constrained by dominant categories (e.g., excessively large water bodies (Cui et al., 2023)) and should be supplemented with additional metrics for a comprehensive evaluation.

6.2.3 Producer accuracy (PA) and user accuracy (UA)

Producer Accuracy (PA) measures, for each reference class, the proportion of samples that are correctly classified, thereby reflecting omission error. User Accuracy (UA) measures, for each predicted class, the proportion of predicted samples that are correct, thereby reflecting commission error and map reliability. For a more balanced evaluation of the classification performance of an individual class, the F1 score (F1) is widely used (Carlson et al., 2024; Khan et al., 2025). When dealing with class-imbalanced datasets, the macro-averaged F1 score (Macro-F1) treats all classes equally, preventing dominant majority classes from driving the evaluation results, and is therefore more suitable for capturing the diversity of wetland vegetation communities (Farhadpour et al., 2024).

Beyond the classic metrics discussed above, a range of more targeted evaluation metrics has been increasingly adopted and actively discussed in recent years (Yu et al., 2025). Table 3 below systematically compares conventional and emerging metrics and summarizes their ecological relevance in wetland studies. Accuracy assessment has evolved into a multi-level metric system grounded in the confusion matrix: overall accuracy and the Kappa coefficient provide baseline evaluation; PA, UA, and the F1 score characterize inter-class differences; IoU/mIoU specifically assess spatial segmentation quality; and significance tests are further applied to ensure the reliability of the conclusions.

7 Conclusion and discussion

7.1 Conclusion

This paper identifies four major developments in remote sensing–based detailed wetland classification over the past 5 years. First, the integration of multisource remote sensing data, including high-resolution optical imagery, multispectral and hyperspectral imagery, SAR, and LiDAR, has become increasingly important for overcoming cloud contamination, spectral confusion, and boundary ambiguity. Second, wetland classification systems have evolved from broad wetland categories toward more detailed subclasses tailored to regional ecological characteristics and management needs. Third, classification methods have progressed from manual feature engineering and traditional machine learning approaches, such as Random Forest and SVM, to deep learning architectures, including CNNs, U-Net, and Transformers, which provide stronger capabilities for extracting complex spatial and spectral patterns. Fourth, accuracy assessment has expanded beyond OA and Kappa coefficients to include class-specific metrics, such as Producer’s Accuracy, User’s Accuracy, F1 score, and uncertainty-aware evaluation, reflecting growing attention to the reliability of detailed wetland mapping. The field has shifted from an initial focus on identifying wetland types and producing maps to greater attention to internal ecological processes and to the need for classification products that can help interpret the driving mechanisms and causes of change.

7.2 Discussion

Despite significant progress in wetland remote sensing classification over the past five years—including advancements in data source utilization, feature engineering, classification methods, and accuracy evaluation—the field still faces numerous challenges and areas requiring in-depth exploration. Future wetland classification research can be further expanded and deepened in the following aspects.

7.2.1 Deep integration and collaborative use of multimodal remote sensing data

The deep integration of multimodal, multiscale, and multi-temporal remote sensing data is emerging as a key pathway to enhance wetland classification accuracy. Future research will increasingly emphasize the synergistic use of data from panchromatic, multispectral, hyperspectral, SAR, and LiDAR sensors to simultaneously capture wetland spatial structure, spectral characteristics, polarization response, and vertical structure information. Leveraging the complementary advantages of high spatial resolution and rich spectral information will enable high-precision identification and boundary extraction of complex wetland types (Jamali et al., 2021a; Jamali et al., 2021b; Xiong et al., 2022; Marjani et al., 2024; Yan et al., 2019; Jamali et al., 2023; Jamali et al., 2022b). Integrating image processing algorithms, such as edge detection, can significantly enhance the ability to distinguish between wetland and land transition zones.

Fusing high-temporal-resolution imagery (e.g., MODIS, Sentinel series) with hyperspectral data will support long-term dynamic monitoring of wetlands by capturing seasonal changes and vegetation phenological characteristics, thereby assisting wetland surveillance. Radar imagery, characterized by its weather-independent penetration capability, effectively compensates for data gaps during cloudy or rainy conditions when fused with optical remote sensing data (Guo et al., 2023). However, existing approaches still underexploit the joint potential of SAR polarimetric information, hyperspectral narrow-band absorption features, and LiDAR-derived vertical structure.

Future efforts should therefore develop cross-modal self-supervised learning frameworks to build an integrated optical, SAR, hyperspectral and LiDAR wetland cube conceptual model, enabling modalities to mutually reinforce each other—for example, using SAR sensitivity to hydrological conditions to constrain vegetation boundary delineation in optical imagery, and leveraging fine hyperspectral signatures to guide the interpretation of SAR polarimetric features.

7.2.2 Intelligent classification methods and system development for wetland scenarios

Current wetland classification research has shifted from traditional machine learning methods toward a new paradigm of multi-algorithm collaboration and deep learning integration. Mainstream deep learning models, such as CNNs, Transformers, and U-Nets, face challenges, including poor interpretability, high reliance on annotations, and limited generalization, making them ill-suited to the complex landforms and dynamic habitats of wetlands. In addition, model complexity and computational cost remain significant challenges. Future research should optimize and adopt high-performance network structures (Radman et al., 2024; Marjani et al., 2024) and explore incorporating wetland ecological knowledge into model design—for example, using vegetation phenology curves as constraints for temporal modeling, hydrological cycles as contextual information for classification decisions, and topographic elevation as prior knowledge for spatial features and concurrently, developing lightweight models with low parameters and generative model (Jamali et al., 2022a; Jamali et al., 2022b) for reducing computational costs and meeting the demands of large-scale wetland monitoring and real-time processing.

One of the major bottlenecks in wetland classification is the scarcity of labeled samples, especially for rare vegetation types and complex transition zones. Therefore, future work should explore learning paradigms such as semi-supervised, self-supervised, domain-adaptive transfer learning, and few-shot learning. Domain adaptation technology effectively mitigates reliance on manually labeled data by narrowing feature distributions between source and target domains, enhancing model adaptability to unlabeled or heterogeneous wetland datasets. This approach collectively improves classification robustness under weakly supervised and small-sample conditions.

Concurrently, wetland classification is moving toward more automated and scalable workflows. Future research should emphasize integrating multisource remote sensing data, advanced classification models, and cloud computing platforms to support large-area mapping and long-term monitoring. In addition, improving the interoperability of data-processing pipelines and classification frameworks will facilitate more efficient wetland inventory updates and dynamic change detection. These developments are expected to enhance the operational applicability of wetland classification products for ecological monitoring and management.

7.2.3 Standardization of wetland classification systems and deepening ecological applications

Resolving inconsistencies in wetland classification systems and scales is a critical prerequisite for enhancing the universality and comparability of research outcomes. China’s current wetland classification system has clear advantages for national inventory and management reporting, yet it has limited interoperability with major international schemes. Future efforts should prioritize interoperability research, strike a balance between scientific rigor and practical applicability, and develop a hierarchical, mappable standard framework that aligns with both international conventions and national standards. Simultaneously, strengthen the application of results, from static mapping to dynamic process simulation, by integrating long-term remote sensing data with landscape ecological indices, ultimately supporting wetland carbon sink estimation, biodiversity conservation, restoration effectiveness assessment, and decision-making for sustainable management in wetland ecosystems.

References

• 1

  AdamE.MutangaO.RugegeD. (2010). Multispectral and hyperspectral remote sensing for identification and mapping of wetland vegetation: a review. Wetl. Ecol. Manag.18, 281–296. 10.1007/s11273-009-9169-z

  • CrossRef
  • Google Scholar

• 2

  AdministrationN. F. a.G. (2009). GB/T 24708-2009 Wetland Classification.

  • Google Scholar

• 3

  AmaniM.GhorbanianA.AhmadiS. A.KakooeiM.MoghimiA.MirmazloumiS. M.et al (2020). Google earth engine cloud computing platform for remote sensing big data applications: a comprehensive review. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens.13, 5326–5350. 10.1109/jstars.2020.3021052

  • CrossRef
  • Google Scholar

• 4

  BelgiuM.DrăguţL. (2016). Random forest in remote sensing: a review of applications and future directions. ISPRS J. Photogrammetry Remote Sensing114, 24–31. 10.1016/j.isprsjprs.2016.01.011

  • CrossRef
  • Google Scholar

• 5

  BhattP.MacleanA. L. (2023). Comparison of high-resolution NAIP and uncrewed aerial vehicle (UAV) imagery for natural vegetation communities classification using machine learning approaches. GIScience Remote Sens.60, 2177448. 10.1080/15481603.2023.2177448

  • CrossRef
  • Google Scholar

• 6

  CaiY. (2021). Research on Spatiotemporal Variation of Vegetation in the Dongting Lake Wetland. Central South University of Forestry and Technology.

  • Google Scholar

• 7

  CarlsonK.ButtenfieldB. P.QiangY. (2024). Wetland classification, attribute accuracy, and scale. ISPRS Int. J. Geo-Information13 (3), 103. 10.3390/ijgi13030103

  • CrossRef
  • Google Scholar

• 8

  ChenL. (2024). Extraction method of alpine wetland information using landsat data. Remote Sens. Technol. Appl.39, 708–717.

  • Google Scholar

• 9

  ChenC. (2006). CiteSpace II: detecting and visualizing emerging trends and transient patterns in scientific literature. J. Am. Soc. Information Sci. Technol.57 (3), 359–377. 10.1002/asi.20317

  • CrossRef
  • Google Scholar

• 10

  ChenC.MaY.RenG.WangJ. (2022). Aboveground biomass of salt-marsh vegetation in coastal wetlands: sample expansion of in situ hyperspectral and Sentinel-2 data using a generative adversarial network. Remote Sens. Environ.270, 112885. 10.1016/j.rse.2021.112885

  • CrossRef
  • Google Scholar

• 11

  ChenZ.ZhangM.ZhangH.LiuY. (2023). Mapping mangrove using a red-edge mangrove index (REMI) based on Sentinel-2 multispectral images. IEEE Trans. Geoscience Remote Sens.61, 1–11. 10.1109/tgrs.2023.3323741

  • CrossRef
  • Google Scholar

• 12

  ChenX.ZhangY.WangS.ZhaoZ.LiuC.WenJ. (2024). Comparative study of machine learning methods for mapping forest fire areas using Sentinel-1B and 2A imagery. Front. Remote Sens.5, 1446641. 10.3389/frsen.2024.1446641

  • CrossRef
  • Google Scholar

• 13

  ChengL. (2022). Rapid and automatic classification of intertidal wetlands based on intensive time series Sentinel-2 images and Google Earth engine. Natl. Remote Sens. Bull.26, 348–357.

  • Google Scholar

• 14

  CongaltonR. G. (1991). A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing Environment37 (1), 35–46. 10.1016/0034-4257(91)90048-b

  • CrossRef
  • Google Scholar

• 15

  CuiB.WuJ.LiX.RenG.LuY. (2023). A combination of deep learning and vegetation index for coastal wetland mapping using GF-2 remote sensing images. Natl. Remote Sens. Bull.27, 1376–1386. 10.11834/jrs.20221658

  • CrossRef
  • Google Scholar

• 16

  DemarquetQ.RapinelS.GoreO.DufourS.Hubert-MoyL. (2024). Continuous change detection outperforms traditional post-classification change detection for long-term monitoring of wetlands. Int. J. Appl. Earth Observation Geoinformation133, 104142. 10.1016/j.jag.2024.104142

  • CrossRef
  • Google Scholar

• 17

  DengY.JiangW.WangX. (2023). Refined wetland classification of international wetland cities based on the random forest algorithm and knowledge-driven rules: a case study of Changde city, China. Natl. Remote Sens. Bull.27, 1426–1440.

  • Google Scholar

• 18

  DuB.MaoD.WangZ.QiuZ.YanH.FengK.et al (2021). Mapping wetland plant communities using unmanned aerial vehicle hyperspectral imagery by comparing object/pixel-based classifications combining multiple machine-learning algorithms. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens.14, 8249–8258. 10.1109/jstars.2021.3100923

  • CrossRef
  • Google Scholar

• 19

  EffahD.ZiaA.AwrangjebM.GaoY.SarpongK. (2025). Advances in machine learning for wetland classification: a comprehensive survey of methods and applications. Artif. Intell. Rev.59 (1), 24. 10.1007/s10462-025-11413-5

  • CrossRef
  • Google Scholar

• 20

  FarhadpourS.WarnerT. A.MaxwellA. E. (2024). Selecting and interpreting multiclass loss and accuracy assessment metrics for classifications with class imbalance: guidance and best practices. Remote Sens.16 (3), 533. 10.3390/rs16030533

  • CrossRef
  • Google Scholar

• 21

  FengQ. (2023). Spectral Characteristic Analysis and Classification of Nanji Wetland Based on Hyperspectral Image. Jiangxi University of Science and Technology.

  • Google Scholar

• 22

  FoodyG. M. (2002). Status of land cover classification accuracy assessment. Remote Sensing Environment80 (1), 185–201. 10.1016/s0034-4257(01)00295-4

  • CrossRef
  • Google Scholar

• 23

  FuB.ZuoP.LiuM.LanG.HeH.LaoZ.et al (2022). Classifying vegetation communities in karst wetlands using synergistic image fusion and object-based machine learning algorithms with Jilin-1 and UAV multispectral images. Ecol. Indic.140, 108989. 10.1016/j.ecolind.2022.108989

  • CrossRef
  • Google Scholar

• 24

  GallantA. L. (2015). The Challenges of Remote Monitoring of Wetlands. MDPI, 10938–10950.

  • Google Scholar

• 25

  GaoY.LiW.ZhangM.WangJ.SunW.TaoR.et al (2021). Hyperspectral and multispectral classification for coastal wetland using depthwise feature interaction network. IEEE Trans. Geoscience Remote Sens.60, 1–15. 10.1109/tgrs.2021.3097093

  • CrossRef
  • Google Scholar

• 26

  GaoY.HuZ.WangZ. (2024). Vegetation classification of Lianyungang coastal wetland based on GF-2 satellite images. Surv. Spatial Geogr. Inf.47, 40–44. 10.3969/j.issn.1672-5867.2024.12.011

  • CrossRef
  • Google Scholar

• 27

  GorelickN.HancherM.DixonM.IlyushchenkoS.ThauD.MooreR. (2017). Google Earth engine: planetary-scale geospatial analysis for everyone. Remote Sensing Environ.202, 18–27. 10.1016/j.rse.2017.06.031

  • CrossRef
  • Google Scholar

• 28

  GouC.PangM.YangY. (2025). Improved U-Net convolutional network application for land cover classification in remote sensing images. Bull. Surv. Mapp. (3), 150–155.

  • Google Scholar

• 29

  GuoY.SuX.LiuW. (2022). Fusion classification of HSI and MSI using a spatial-spectral vision transformer for wetland biodiversity estimation. Remote Sens.14, 850.

  • Google Scholar

• 30

  GuoF.LiZ.MengQ.RenG.WangL.WangJ.et al (2023). Semi-supervised cross-domain feature fusion classification network for coastal wetland classification with hyperspectral and LiDAR data. Int. J. Appl. Earth Observation Geoinformation120, 103354. 10.1016/j.jag.2023.103354

  • CrossRef
  • Google Scholar

• 31

  HanY.KeY.WangZ.LiangD.ZhouD. (2023). Classification of the Yellow River Delta wetland landscape based on ZY-1 02D hyperspectral imagery. Natl. Remote Sens. Bull.27, 1387–1399. 10.11834/jrs.20211071

  • CrossRef
  • Google Scholar

• 32

  HosseinyB.MahdianpariM.BriscoB.MohammadimaneshF.SalehiB. (2022). WetNet: a spatial-temporal ensemble deep learning model for wetland classification using Sentinel-1 and Sentinel-2. IEEE Trans. Geoscience Remote Sens.60, 4406014–14. 10.1109/tgrs.2021.3113856

  • CrossRef
  • Google Scholar

• 33

  HuQ.WoldtW.NealeC.ZhouY.DrahotaJ.VarnerD.et al (2021). Utilizing unsupervised learning, multi-view imaging, and CNN-based attention facilitates cost-effective wetland mapping. Remote Sens. Environ.267, 112757. 10.1016/j.rse.2021.112757

  • CrossRef
  • Google Scholar

• 34

  HuangY.PengJ.ChenN.SunW.DuQ.RenK.et al (2023a). Cross-scene wetland mapping on hyperspectral remote sensing images using adversarial domain adaptation network. ISPRS J. Photogrammetry Remote Sens.203, 37–54. 10.1016/j.isprsjprs.2023.07.009

  • CrossRef
  • Google Scholar

• 35

  HuangY.LuC.JiaM.WangZ.SuY.SuY. (2023b). Plant species classification of coastal wetlands based on UAV images and object-oriented deep learning. Biodivers. Sci.31, 143–158. 10.17520/biods.2022411

  • CrossRef
  • Google Scholar

• 36

  HuangY.YangG.SunW.ZhuL.HuangK.MengX. (2023c). Wetlands mapping in typical regions of South America with multisource and multi-feature integration. Natl. Remote Sens. Bull.27, 1300–1319. 10.11834/jrs.20232265

  • CrossRef
  • Google Scholar

• 37

  HuangB.XuJ.ZhangX. (2024). Wetland information extraction from GF-6 WFV multispectral imagery based on feature optimization. Geospatial Inf.22, 39–44.

  • Google Scholar

• 38

  HuoX.NiuZ. (2024). Fine-resolution wetland mapping in the Yellow River Basin using Sentinel-1/2 data via zoning-based random forest with remote sensing feature preferences. Water16 (17), 2415. 10.3390/w16172415

  • CrossRef
  • Google Scholar

• 39

  IgweV.SalehiB.MahdianpariM. (2023). Rapid large-scale wetland inventory update using multisource remote sensing. Remote Sens.15 (20), 4960. 10.3390/rs15204960

  • CrossRef
  • Google Scholar

• 40

  JafarzadehH.MahdianpariM.GillE. W. (2022). Wet-GC: a novel multimodel graph convolutional approach for wetland classification using Sentinel-1 and 2 imagery with limited training samples. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens.15, 5303–5316. 10.1109/jstars.2022.3177579

  • CrossRef
  • Google Scholar

• 41

  JamaliA.MahdianpariM.BriscoB.GrangerJ.MohammadimaneshF.SalehiB. (2021a). Comparing solo versus ensemble convolutional neural networks for wetland classification using multispectral satellite imagery. Remote Sens.13, 2046. 10.3390/rs13112046

  • CrossRef
  • Google Scholar

• 42

  JamaliA.MahdianpariM.MohammadimaneshF.BriscoB.SalehiB. (2021b). A synergic use of Sentinel-1 and Sentinel-2 imagery for complex wetland classification using generative adversarial network (GAN) scheme. Water13, 3601. 10.3390/w13243601

  • CrossRef
  • Google Scholar

• 43

  JamaliA.MahdianpariM.MohammadimaneshF.HomayouniS. (2022a). A deep learning framework based on generative adversarial networks and vision transformer for complex wetland classification using limited training samples. Int. J. Appl. Earth Observation Geoinformation115, 103095. 10.1016/j.jag.2022.103095

  • CrossRef
  • Google Scholar

• 44

  JamaliA.MahdianpariM.BriscoB.MaoD.SalehiB.MohammadimaneshF. (2022b). 3DUNetGSFormer: a deep learning pipeline for complex wetland mapping using generative adversarial networks and swin transformer. Ecol. Inf.72, 101904. 10.1016/j.ecoinf.2022.101904

  • CrossRef
  • Google Scholar

• 45

  JamaliA.KumarS. R.GhamisiP. (2023). WetMapFormer: a unified deep CNN and vision transformer for complex wetland mapping. Int. J. Appl. Earth Observation Geoinformation120, 103333. 10.1016/j.jag.2023.103333

  • CrossRef
  • Google Scholar

• 46

  JarocińskaA.KopećD.KyckoM. (2024). Comparison of dimensionality reduction methods on hyperspectral images for the identification of heathlands and mires. Sci. Rep.14 (1), 27662. 10.1038/s41598-024-79209-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  JiangR.LiuJ.LiuW.ZhangD.HuW. (2023). Changes and driving forces analysis of alpine wetlands in the first meander of the yellow river based on long-term time series remote sensing data. Front. Ecol. Evol.11, 1193059. 10.3389/fevo.2023.1193059

  • CrossRef
  • Google Scholar

• 48

  JudahA.HuB. (2022). The integration of multisource remotely sensed data with hierarchically based classification approaches in support of the classification of wetlands. Can. J. Remote Sens.48, 158–181. 10.1080/07038992.2021.1967732

  • CrossRef
  • Google Scholar

• 49

  KhanM. S. I.Vega-CorredorM. C.WilsonM. D. (2025). Mapping wetlands with high-resolution planet SuperDove satellite imagery: an assessment of machine learning models across the diverse waterscapes of New Zealand. Remote Sens.17 (15), 2626. 10.3390/rs17152626

  • CrossRef
  • Google Scholar

• 50

  Laporte-FauretQ.LubacB.CastelleB.MichaletR.MarieuV.BombrunL.et al (2020). Classification of Atlantic coastal sand dune vegetation using in situ,, UAV, and airborne hyperspectral data. Remote Sens.12, 2222. 10.3390/rs12142222

  • CrossRef
  • Google Scholar

• 51

  LeishiC. (2021). Wetland classification method of the Dongting Lake district based on CART using GF-2 image. Bull. Surv. Mapp. (6), 12–15.

  • Google Scholar

• 52

  LiY.YuX.ZhangJ.ZhangS.WangX.KongD.et al (2024a). Improved classification of coastal wetlands in the Yellow River Delta of China using ResNet combined with feature-preferred bands based on attention mechanism. Remote Sens.16, 1860. 10.3390/rs16111860

  • CrossRef
  • Google Scholar

• 53

  LiK.YueL.TanQ. (2024). Precise mapping of coastal wetlands using time-series remote sensing images and a deep learning model. Front. For. Glob. Change7, 1409985.

  • Google Scholar

• 54

  LiW.FanY.ZhouP. (2024). Improved transformer-based hyperspectral image classification method for surface features: a case study of the Yellow River Delta. Remote Sens. Land Resour.36, 137–145.

  • Google Scholar

• 55

  LiY.LiY.ZuoX. (2024b). A polarimetric SAR image ground recognition method based on U-Net network combined with multiple attention mechanisms. Sci. Surv. Mapp.49, 69–76. 10.16251/j.cnki.1009-2307.2024.10.007

  • CrossRef
  • Google Scholar

• 56

  LiZ.WangZ.PanJ.JiaoX.FanY.GaoZ.et al (2025a). Exploring the potential of time series InSAR for coastal wetland fine mapping using auto-weighted ensemble machine learning. Int. J. Digital Earth18 (1), 2528619. 10.1080/17538947.2025.2528619

  • CrossRef
  • Google Scholar

• 57

  LiZ.LiuT.LuY.TianJ.ZhangM.ZhouC. (2025b). Enhanced hyperspectral image classification for coastal wetlands using a hybrid CNN-transformer approach with cross-attention mechanism. Front. Mar. Sci.12, 1613565. 10.3389/fmars.2025.1613565

  • CrossRef
  • Google Scholar

• 58

  LijiaC. (2025). CATrans: a cross-scale attention transformer for land cover semantic segmentation in high-resolution remote sensing images. J. Geo-information Sci.27, 1624–1637.

  • Google Scholar

• 59

  LinY.GuoJ. (2024). Fuzzy geospatial objects-based wetland remote sensing image classification: a case study of Tianjin Binhai new area. Int. J. Appl. Earth Observation Geoinformation132, 104051. 10.1016/j.jag.2024.104051

  • CrossRef
  • Google Scholar

• 60

  LiuX. (2023). Research on Wetland Vegetation Classification Based on Generative Adversarial Learning. Master’s thesis. Chengdu: Chengdu University of Technology. 10.26986/d.cnki.gcdlc.2023.001603

  • CrossRef
  • Google Scholar

• 61

  LiuR.LiJ.SunC. (2021). Classification of Yancheng coastal wetland vegetation based on vegetation phenological characteristics derived from Sentinel-2 time-series. Acta Geogr. Sin.76, 1680–1692.

  • Google Scholar

• 62

  LiuK.SunW.ShaoY.LiuW.YangG.MengX.et al (2022). Mapping coastal wetlands using transformer in transformer deep network on China ZY1-02D hyperspectral satellite images. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens.15, 3891–3903. 10.1109/jstars.2022.3173349

  • CrossRef
  • Google Scholar

• 63

  LiuH.JiangQ.MaY.YangQ.ShiP.ZhangS.et al (2022). Object-based multigrained cascade forest method for wetland classification using sentinel-2 and radarsat-2 imagery. Water14, 82. 10.3390/w14010082

  • CrossRef
  • Google Scholar

• 64

  LiuY.LiuK.CaoJ. (2023). Classification of coastal wetlands in the Pearl river Estuary using Zhuhai-1 hyperspectral imagery and XGBoost algorithm. Bull. Surv. Mapp. (12), 136–141.

  • Google Scholar

• 65

  LiuW.LiuK.SunW.YangG.RenK.MengX.et al (2023). Self-supervised feature learning based on spectral masking for hyperspectral image classification. IEEE Trans. Geoscience Remote Sens.61, 1–15. 10.1109/tgrs.2023.3310489

  • CrossRef
  • Google Scholar

• 66

  LiuH.BiW.MugheesN. (2025). Enhanced hyperspectral image classification technique using PCA-2D-CNN algorithm and null spectrum hyperpixel features. Sensors25 (18), 5790. 10.3390/s25185790

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 67

  LiuY.LiuQ.ZhangX. (2025). Classification of vegetation information by integrating color features with multi-feature optimization of random forest in the Yellow River Delta. Geomatics Inf. Sci. Wuhan Univ.50, 1802–1811.

  • Google Scholar

• 68

  LongT.KuangR.FengQ. (2025). Classification of the Nanji Wetland based on Zhuhai-1 hyperspectral imagery. Wetl. Sci.23, 867–877.

  • Google Scholar

• 69

  LouA.HeZ.XiaoM. (2024). Wetland classification based on Sentinel-2 and 3D multisource domain self-attention model. Natl. Remote Sens. Bull.28, 266–279. 10.11834/jrs.20232246

  • CrossRef
  • Google Scholar

• 70

  LuoB.MaoD.ZhangJ. (2025). Fine classification of urban wetlands based on Sentinel-1/2 images and random forest classification in the built-up area of Haikou. Sci. Geogr. Sin.45, 389–402.

  • Google Scholar

• 71

  LuoK.SamatA.AbuduwailiJ.LiW. (2023). Evaluation of remote sensing products for wetland mapping in the Irtysh River Basin. Geosciences14 (1), 14. 10.3390/geosciences14010014

  • CrossRef
  • Google Scholar

• 72

  LuoK.SamatA.Van de voordeT.JiangW.LiW.AbuduwailiJ. (2025). An automatic classification method with weak supervision for large-scale wetland mapping in transboundary (Irtysh River) Basin using sentinel 1/2 imageries. J. Environ. Manag.380, 124969. 10.1016/j.jenvman.2025.124969

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 73

  MaH.FengT.JiR. (2022). Land cover classification of high-resolution remote sensing imagery based on FCN and object-oriented technology. Hubei Agric. Sci.61, 163–168.

  • Google Scholar

• 74

  MaS.YuY.WangC. (2023). Extraction of typical features based on improved U-Net. Geomatics and Spatial Inf. Technol.46, 82–85.

  • Google Scholar

• 75

  MahdaviS.SalehiB.GrangerJ.AmaniM.BriscoB.HuangW. (2018). Remote sensing for wetland classification: a comprehensive review. GIScience and Remote Sensing55 (5), 623–658. 10.1080/15481603.2017.1419602

  • CrossRef
  • Google Scholar

• 76

  MahdianpariM.SalehiB.MohammadimaneshF.HomayouniS.GillE. (2018). The first wetland inventory map of newfoundland at a spatial resolution of 10 m using sentinel-1 and sentinel-2 data on the google earth engine cloud computing platform. Remote Sensing11 (1), 43. 10.3390/rs11010043

  • CrossRef
  • Google Scholar

• 77

  MahdianpariM.BriscoB.GrangerJ. E.MohammadimaneshF.SalehiB.BanksS.et al (2020a). The second generation Canadian wetland inventory map at 10 meters resolution using Google Earth Engine. Can. J. Remote Sens.46 (3), 360–375. 10.1080/07038992.2020.1802584

  • CrossRef
  • Google Scholar

• 78

  MahdianpariM.GrangerJ. E.MohammadimaneshF.SalehiB.BriscoB.HomayouniS.et al (2020b). Meta-analysis of wetland classification using remote sensing: a systematic review of a 40-year trend in North America. Remote Sens.12 (11), 1882. 10.3390/rs12111882

  • CrossRef
  • Google Scholar

• 79

  MaoD.WangZ.DuB.LiL.TianY.JiaM.et al (2020). National wetland mapping in China: a new product resulting from object-based and hierarchical classification of Landsat 8 OLI images. ISPRS J. Photogrammetry Remote Sens.164, 11–25. 10.1016/j.isprsjprs.2020.03.020

  • CrossRef
  • Google Scholar

• 80

  MarjaniM.MahdianpariM.MohammadimaneshF.GillE. W. (2024). CVTNet: a fusion of convolutional neural networks and vision transformer for wetland mapping using Sentinel-1 and Sentinel-2 satellite data. Remote Sens.16, 2427. 10.3390/rs16132427

  • CrossRef
  • Google Scholar

• 81

  MartinsS. V.KaleitaL. A.GelderK. B.NagelG. W.MacielD. A. (2020). Deep neural network for complex open-water wetland mapping using high-resolution WorldView-3 and airborne LiDAR data. Int. J. Appl. Earth Observation Geoinformation93, 102215. 10.1016/j.jag.2020.102215

  • CrossRef
  • Google Scholar

• 82

  MengX. (2019). Study on Remote Sensing Finer Classification of Freshwater Wetland Based on Deep Learning. Master’s thesis. Harbin: Harbin Normal University. 10.27064/d.cnki.ghasu.2019.000087

  • CrossRef
  • Google Scholar

• 83

  MingY.LiuQ.BaiH.HuangC. (2023). Classification and change detection of vegetation in the Ruoergai Wetland using optical and SAR remote sensing data. Natl. Remote Sens. Bull.27, 1414–1425. 10.11834/jrs.20221767

  • CrossRef
  • Google Scholar

• 84

  NaX.LiX.LiW.WuC. (2021). Wetland mapping using HJ-1A/B hyperspectral images and an adaptive sparse constrained least squares linear spectral mixture model. Remote Sens.13 (4), 751. 10.3390/rs13040751

  • CrossRef
  • Google Scholar

• 85

  NationsU. (2015). Transforming our world: the 2030 agenda for sustainable development. Available online at: https://sdgs.un.org/2030agenda (Accessed December 23, 2025).

  • Google Scholar

• 86

  O'NeilG. L.GoodallJ. L.BehlM.SabyL. (2020). Deep learning using physically-informed input data for wetland identification. Environ. Model. and Softw.126, 104665. 10.1016/j.envsoft.2020.104665

  • CrossRef
  • Google Scholar

• 87

  PanY.LinH. (2022). Classification of the Sanjiang Wetland in Heilongjiang province based on domestic high-resolution satellite data. J. Central South Univ. For. Technol.42, 71–81.

  • Google Scholar

• 88

  PanC.HouH.TangW. (2024). Refined classification of the wetlands in Hangzhou by Google Earth Engine (GEE) and Sentinel-2 imagery. J. Beijing Normal Univ. Nat. Sci.60, 447–458.

  • Google Scholar

• 89

  PengK.JiangW.HouP.WuZ.CuiT. (2023). Detailed wetland-type classification using Landsat-8 time-series images: a pixel- and object-based algorithm with knowledge (POK). GIScience and Remote Sens.61, 1–20. 10.1080/15481603.2023.2293525

  • CrossRef
  • Google Scholar

• 90

  QianH.BaoN.MengD.ZhouB.LeiH.LiH. (2024). Mapping and classification of Liao River Delta coastal wetland based on time series and multisource GaoFen images using a stacking ensemble model. Ecol. Inf.80, 102488. 10.1016/j.ecoinf.2024.102488

  • CrossRef
  • Google Scholar

• 91

  QianS.XueZ.JiaM.ChenY.SuH. (2024). Temporal-spectral-semantic-aware convolutional transformer network for multi-class tidal wetland change detection in Greater Bay Area. ISPRS J. Photogrammetry Remote Sens.216, 126–141. 10.1016/j.isprsjprs.2024.07.024

  • CrossRef
  • Google Scholar

• 92

  RadmanA.MohammadimaneshF.MahdianpariM. (2024). Wet-ConViT: a hybrid convolutional-transformer model for efficient wetland classification using satellite data. Remote Sens.16, 2673. 10.3390/rs16142673

  • CrossRef
  • Google Scholar

• 93

  Ramsar (2024). Ramsar Wetland Classification System. Available online at: https://www.ramsar.org/document/sc63-doc19-annex-1-ramsar-wetland-classification-system (Accessed November 12, 2025).

  • Google Scholar

• 94

  RenK.SunW.MengX.YangG.DuQ. (2020). Fusing China GF-5 hyperspectral data with GF-1, GF-2 and Sentinel-2A multispectral data: which methods should be used?Remote Sens.12 (5), 882. 10.3390/rs12050882

  • CrossRef
  • Google Scholar

• 95

  RonnebergerO.FischerP.BroxT. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation.

  • Google Scholar

• 96

  SahourH.KeminkK. M.O’ConnellJ. (2021). Integrating SAR and optical remote sensing for conservation-targeted wetlands mapping. Remote Sens.14, 159. 10.3390/rs14010159

  • CrossRef
  • Google Scholar

• 97

  ServiceU. S. F. A.W. (2023). Wetland classification codes. Available online at: https://www.fws.gov/program/national-wetlands-inventory/classification-codes (Accessed November 12, 2025).

  • Google Scholar

• 98

  ShiT.WangC.ZhangW.HeJ. (2025). Classification of coastal wetlands in the Liaohe Delta with multisource and multi-feature integration. Environ. Monit. Assess.197, 247. 10.1007/s10661-025-13717-z

  • CrossRef
  • Google Scholar

• 99

  ShuF.CaoJ.WuQ. (2024). “Patch-based cascade forest wetland classification based on multi-temporal SAR images in Yellow River Delta,” in 2024 IEEE International Geoscience and Remote Sensing Symposium (IGARSS) (Athens, Greece: IEEE).

  • Google Scholar

• 100

  SunW. W. (2023). Research progress and prospects of hyperspectral remote sensing for global wetland from 2010 to 2022. Natl. Remote Sens. Bull.27 (6), 1281–1299.

  • Google Scholar

• 101

  SunZ. (2024). Rapid expansion of coastal mangrove forest in Guangxi Beibu Gulf: patterns, drivers and impacts. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens. 18, 510–522.

  • Google Scholar

• 102

  SunS.ZhangY.SongZ.ChenB.YuanW.et al (2020). Mapping coastal wetlands of the Bohai Rim at a spatial resolution of 10 m using multiple open-access satellite data and terrain indices. Remote Sens.12, 4114. 10.3390/rs12244114

  • CrossRef
  • Google Scholar

• 103

  SunW.LiuK.RenG.LiuW.YangG.MengX.et al (2021). A simple and effective spectral-spatial method for mapping large-scale coastal wetlands using China ZY1-02D satellite hyperspectral images. Int. J. Appl. Earth Observation Geoinformation104, 102572. 10.1016/j.jag.2021.102572

  • CrossRef
  • Google Scholar

• 104

  SunW.ChenD.LiZ.LiS.ChengS.NiuX.et al (2024). Monitoring wetland plant diversity from space: progress and perspective. Int. J. Appl. Earth Observation Geoinformation130, 103943. 10.1016/j.jag.2024.103943

  • CrossRef
  • Google Scholar

• 105

  TamiminiaH.SalehiB.MahdianpariM.QuackenbushL.AdeliS.BriscoB. (2020). Google Earth Engine for geo-big data applications: a meta-analysis and systematic review. ISPRS Journal Photogrammetry Remote Sensing164, 152–170. 10.1016/j.isprsjprs.2020.04.001

  • CrossRef
  • Google Scholar

• 106

  TangX.DouZ.ZhaoX. (2025). UAV hyperspectral mapping of Spartina alterniflora in Yancheng coastal wetland restoration area. Ecol. Sci.44, 108–118. 10.14108/j.cnki.1008-8873.2025.01.012

  • CrossRef
  • Google Scholar

• 107

  TianY.ChenZ.HuiF. (2019). ESA Sentinel-2A/B satellite: characteristics and applications. J. Beijing Normal Univ. Nat. Sci.55, 57–65.

  • Google Scholar

• 108

  TootchiA.JostA.DucharneA. (2019). Multisource global wetland maps combining surface water imagery and groundwater constraints. Earth Syst. Sci. Data11, 189–220. 10.5194/essd-11-189-2019

  • CrossRef
  • Google Scholar

• 109

  TuC.LiP.LiZ.WangH.YinS.LiD.et al (2021). Synergetic classification of coastal wetlands over the Yellow River Delta with GF-3 full-polarization SAR and Zhuhai-1 OHS hyperspectral remote sensing. Remote Sens.13, 4444. 10.3390/rs13214444

  • CrossRef
  • Google Scholar

• 110

  TuiaD.PerselloC.BruzzoneL. (2016). Domain adaptation for the classification of remote sensing data: an overview of recent advances. IEEE Geoscience Remote Sensing Magazine4 (2), 41–57. 10.1109/mgrs.2016.2548504

  • CrossRef
  • Google Scholar

• 111

  TurnbullA.Soto-BerelovM.CooteM. (2024). Delineation and classification of wetlands in the Northern Jarrah Forest, Western Australia using remote sensing and machine learning. Wetlands44, 68. 10.1007/s13157-024-01806-7

  • CrossRef
  • Google Scholar

• 112

  WangM.MaoD.WangY.XiaoX.XiangH.FengK.et al (2023a). Wetland mapping in East Asia by two-stage object-based Random Forest and hierarchical decision tree algorithms on Sentinel-1/2 images. Remote Sens. Environ.297, 113793. 10.1016/j.rse.2023.113793

  • CrossRef
  • Google Scholar

• 113

  WangX.GaoX.ZhangY.FeiX.ChenZ.WangJ.et al (2019). Land-cover classification of coastal wetlands using the RF algorithm for WorldView-2 and Landsat 8 images. Remote Sens.11, 1927. 10.3390/rs11161927

  • CrossRef
  • Google Scholar

• 114

  WangM.HeM.WenN. (2023b). Polarimetric SAR coastal wetland feature classification based on combined optimization features. Geospatial Inf.21, 63–67.

  • Google Scholar

• 115

  WangY.JinS.DardanelliG. (2024). Vegetation classification and evaluation of Yancheng coastal wetlands based on random forest algorithm from Sentinel-2 images. Remote Sens.16, 1124. 10.3390/rs16071124

  • CrossRef
  • Google Scholar

• 116

  WangZ.CaoJ.ShuF. (2025). Wetland mapping method integrating deep features and SVM using Sentinel-1/2. Sci. Surv. Mapp.50, 114–124.

  • Google Scholar

• 117

  WangB.SunY.ZhuX.TengS.LiY. (2025). A modified Swin-UNet model for coastal wetland classification using multi-temporal Sentinel-2 images. Estuaries Coasts48, 918–934. 10.1007/s12237-025-01498-0

  • CrossRef
  • Google Scholar

• 118

  WenL.RyanS.PowellM.LingJ. E. (2025). From clusters to communities: enhancing wetland vegetation mapping using unsupervised and supervised synergy. Remote Sens.17 (13), 2279. 10.3390/rs17132279

  • CrossRef
  • Google Scholar

• 119

  WetlandsR. C. o. (2018). Wetlands and the SDGs: Scaling up Wetland Conservation, Wise Use and Restoration to Achieve the Sustainable Development Goals.

  • Google Scholar

• 120

  WuX.JiangW.PengK. (2022). A framework for fine classification of urban wetlands based on random forest and knowledge rules: taking the wetland cities of Haikou and Yinchuan as examples. GIScience and Remote Sens.59, 2144–2163.

  • Google Scholar

• 121

  WuW.LinZ.ChenC.ChenZ.ZhaoZ.SuH. (2024). Tracking the dynamics of tidal wetlands with time-series satellite images in the Yangtze River Estuary, China. Int. J. Digital Earth17 (1), 2330684. 10.1080/17538947.2024.2330684

  • CrossRef
  • Google Scholar

• 122

  WuY.XuX.LiuY.ShiS.MaA.ZhongY. (2025). Large-scale tidal wetland mapping based on point-sample-driven debiased and label noise learning: a case study in Asia. Int. J. Appl. Earth Observation Geoinformation144, 104913. 10.1016/j.jag.2025.104913

  • CrossRef
  • Google Scholar

• 123

  XinZ.LiZ.WangL. (2023). Hyperspectral image classification of Yellow River Delta wetlands based on a spectral-spatial unified transformer model. Mar. Sci.47, 90–101.

  • Google Scholar

• 124

  XinZ.LiZ.XuM.WangL.RenG.WangJ.et al (2024). Feature disentanglement based domain adaptation network for cross-scene coastal wetland hyperspectral image classification. Int. J. Appl. Earth Observation Geoinformation129, 103850. 10.1016/j.jag.2024.103850

  • CrossRef
  • Google Scholar

• 125

  XiongC.ZhaoZ.CaoJ. (2022). Application of a data set expansion method based on generative adversarial network in wetland bird image recognition. J. Tianjin Univ. Sci. Technol.37, 64–70.

  • Google Scholar

• 126

  XuS.GengS.YangQ.GaoH. (2023). Hyperspectral image dimensionality reduction algorithm based on spatial–spectral adaptive multiple manifolds. Appl. Sci.13 (16), 9180. 10.3390/app13169180

  • CrossRef
  • Google Scholar

• 127

  XuM.LiZ.WangL. (2023). Combined Bi-LSTM and attention enhancement multiscale convolution for wetland hyperspectral image classification in the Yellow River Delta. Mar. Sci.47, 27–40.

  • Google Scholar

• 128

  XueZ.QianS. (2023). Two-stream translating LSTM network for mangroves mapping using Sentinel-2 multivariate time series. IEEE Trans. Geoscience Remote Sens.61, 4401416–16. 10.1109/tgrs.2023.3249179

  • CrossRef
  • Google Scholar

• 129

  XueY.ZhangS.BaiN. (2024). Classification of mangrove species using coupled Sentinel-1/2 multisource time-series remote sensing image features. J. Geo-information Sci.26, 2626–2642.

  • Google Scholar

• 130

  YanJ.ChenH.LiuL. (2019). Overview of hyperspectral image classification. Opt. Precis. Eng.27, 680–693. 10.3788/ope.20192703.0680

  • CrossRef
  • Google Scholar

• 131

  YangX.ZhuZ.QiuS.KroegerK. D.ZhuZ.CovingtonS. (2022). Detection and characterization of coastal tidal wetland change in the northeastern US using Landsat time series. Remote Sens. Environ.276, 113047. 10.1016/j.rse.2022.113047

  • CrossRef
  • Google Scholar

• 132

  YangZ.HuangY.WangY. (2022). Coastal wetland vegetation classification using pixel-based, object-based, and deep learning methods based on RGB-UAV. Land11, 2039.

  • Google Scholar

• 133

  YangR.LuoF.RenF.HuangW.LiQ.DuK.et al (2022). Identifying urban wetlands through remote sensing scene classification using deep learning: a case study of Shenzhen, China. ISPRS Int. J. Geo-Information11, 131. 10.3390/ijgi11020131

  • CrossRef
  • Google Scholar

• 134

  YangM.LiuE.YangY. (2025). Quantitative analysis of wetland evolution characteristics and driving factors in Ruoergai Plateau based on Landsat time series remote sensing images. Environ. Sci. (Huanjing Kexue)46, 4461–4472.

  • Google Scholar

• 135

  YangZ.JiangW.YinX.LingZ.WangX.LinM.et al (2025). A generalizable framework for urban wetland training samples generation and migration: a case study of global Ramsar Wetland Cities. Ecol. Indic.178, 114078. 10.1016/j.ecolind.2025.114078

  • CrossRef
  • Google Scholar

• 136

  YaoH.FuB.ZhangY.LiS.XieS.QinJ.et al (2022). Combination of hyperspectral and quad-polarization SAR images to classify marsh vegetation using stacking ensemble learning algorithm. Remote Sens.14 (21), 5478. 10.3390/rs14215478

  • CrossRef
  • Google Scholar

• 137

  YinX.JiangW.YangZ.LingZ.TsendbazarN. E.HouP.et al (2026). Hybrid wetland city map: improved wetland characterization through the synergy of global land cover products. Int. J. Appl. Earth Observation Geoinformation146, 104994. 10.1016/j.jag.2025.104994

  • CrossRef
  • Google Scholar

• 138

  YuD.RenL.ChenC.KongX.ZhouM.YangL.et al (2025). An AttSDNet model for multiscale feature perception enhanced remote sensing classification of coastal salt-marsh wetlands. Mar. Environ. Res.204, 106899. 10.1016/j.marenvres.2024.106899

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 139

  YueL.WangM.HuangC.ChengQ.YuanQ.ShenH. (2025). Mapping hierarchical wetland characteristics by optical-SAR integration with collaborative spatial-spectral-temporal learning. Int. J. Appl. Earth Observation Geoinformation136, 104395. 10.1016/j.jag.2025.104395

  • CrossRef
  • Google Scholar

• 140

  ZhangY.LiuY.LiuM. (2023). Wetland plant community classification based on Relief F and a convolutional neural network. Bull. Surv. Mapp. (2), 58–64.

  • Google Scholar

• 141

  ZhangJ. (2010). Multisource remote sensing data fusion: status and trends. Int. Journal Image Data Fusion1 (1), 5–24. 10.1080/19479830903561035

  • CrossRef
  • Google Scholar

• 142

  ZhangS.YanF.YuL. (2013). Application of remote sensing technology to wetland research. Sci. Geogr. Sin.33, 1406–1412.

  • Google Scholar

• 143

  ZhangY.YangH.YuanC. (2018). A survey of remote sensing image classification methods. J. Ordnance Equip. Eng.39, 108–112.

  • Google Scholar

• 144

  ZhangL.GongZ.WangQ. (2019). Wetland mapping of Yellow River Delta wetlands based on multi-feature optimization of Sentinel-2 images. Natl. Remote Sens. Bull.23, 313–326. 10.11834/jrs.20198083

  • CrossRef
  • Google Scholar

• 145

  ZhangX.LiuL.ZhaoT.ChenX.LinS.WangJ.et al (2023). GWL_FCS30: a global 30 m wetland map with a fine classification system using multi-sourced and time-series remote sensing imagery in 2020. Earth Syst. Sci. Data15 (1), 265–293. 10.5194/essd-15-265-2023

  • CrossRef
  • Google Scholar

• 146

  ZhangY.FuB.SunX.YaoH.ZhangS.WuY.et al (2023). Effects of multi-growth periods UAV images on classifying karst wetland vegetation communities using object-based optimization stacking algorithm. Remote Sens.15, 4003. 10.3390/rs15164003

  • CrossRef
  • Google Scholar

• 147

  ZhangL.LuoW.ZhangH.YinX.LiB. (2023). Classification scheme for mapping wetland herbaceous plant communities using time series Sentinel-1 and Sentinel-2 data. Natl. Remote Sens. Bull.27, 1362–1375. 10.11834/jrs.20222079

  • CrossRef
  • Google Scholar

• 148

  ZhangX.LiuL.ZhaoT.WangJ.LiuW.ChenX. (2024). Global annual wetland dataset at 30 m with a fine classification system from 2000 to 2022. Sci. Data11 (1), 310. 10.1038/s41597-024-03143-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 149

  ZhangJ.LiuX.QinY.FanY.ChengS. (2024). Wetlands mapping and monitoring with long-term time series satellite data based on Google Earth engine, random forest, and feature optimization: a case study in Gansu Province, China. Land13 (9), 1527. 10.3390/land13091527

  • CrossRef
  • Google Scholar

• 150

  ZhangM.KeY.ShangK.ZhuoZ.LiuH.LiP.et al (2025). Suaeda Salsa Hyperspectral Index (SSHI) for mapping S. salsa in coastal wetlands using hyperspectral satellite imagery: case studies in northern coastal China. GIScience and Remote Sens.62 (1), 2500793. 10.1080/15481603.2025.2500793

  • CrossRef
  • Google Scholar

• 151

  ZhangZ.JiangW.SongJ.LingZ.YangZ.Van de VoordeT.et al (2025). Mapping the first global long time series wetland multitype sample dataset via the Google Earth Engine: a hybrid method of automated generation–index thresholding–spectral matching. GIScience and Remote Sens.62 (1), 2553942. 10.1080/15481603.2025.2553942

  • CrossRef
  • Google Scholar

• 152

  ZhangL.WangZ.WangJ.HuQ.ChangY.LuZ.et al (2025). An integrated feature framework for wetland mapping using multisource imagery. Remote Sens.17 (22), 3737. 10.3390/rs17223737

  • CrossRef
  • Google Scholar

• 153

  ZhaoX.TianB.NiuY. (2022). Classification of coastal salt marsh based on Sentinel-1 time series backscattering characteristics: the case of the Yangtze River Delta. Natl. Remote Sens. Bull.26, 672–682.

  • Google Scholar

• 154

  ZhengJ.SunC.LinY. (2023). Classification of salt marsh vegetation based on pixel-level time series from Landsat images. Natl. Remote Sens. Bull.27, 1400–1413.

  • Google Scholar

• 155

  ZhouN. (2024). ViT-UNet: a vision transformer based UNet model for coastal wetland classification based on high spatial resolution imagery. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens. 17, 19575–19587.

  • Google Scholar

• 156

  ZhouC.ZhaoQ.WuT.LiuX.ChenY. (2024). Distribution and long-term variation of wetland land cover types in the Yellow River Delta remote sensing monitoring. J. Mar. Sci. Eng.12 (12), 2345. 10.3390/jmse12122345

  • CrossRef
  • Google Scholar

• 157

  ZhuX. X.TuiaD.MouL.XiaG. S.ZhangL.XuF.et al (2017). Deep learning in remote sensing: a comprehensive review and list of resources. IEEE Geoscience Remote Sensing Magazine5 (4), 8–36. 10.1109/mgrs.2017.2762307

  • CrossRef
  • Google Scholar

• 158

  ZhuH.HuangY.LiY.YuF.ZhangG.FanL.et al (2022). Predicting plant diversity in beach wetland downstream of Xiaolangdi reservoir with UAV and satellite multispectral images. Sci. Total Environ.819, 153059. 10.1016/j.scitotenv.2022.153059

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 159

  ZhuY.MyintS. W.CaoJ.LiuK.ZengM.DiaoC. (2025). Evaluating multitemporal vegetation indices from Zhuhai-1 hyperspectral images for detecting a rapidly spreading invasive species-Spartina alterniflora. Ecol. Inf.90, 103208. 10.1016/j.ecoinf.2025.103208

  • CrossRef
  • Google Scholar

• 160

  ZhuangX.KuangR. (2024). Object-oriented decision tree classification of the Nanji Wetland based on Sentinel-2 data. Wetl. Sci.22, 630–640.

  • Google Scholar

• 161

  ZongY.LiY.LiuH. (2021). A study of coastal wetland vegetation classification based on object-oriented random forest method. J. Nanjing Normal Univ. Eng. Technol. Ed.21, 47–55.

  • Google Scholar

• 162

  ZouJ.HeW.WangH.ZhangH. (2025). SAM-CTMapper: utilizing segment anything model and scale-aware mixed CNN-transformer facilitates coastal wetland hyperspectral image classification. Int. J. Appl. Earth Observation Geoinformation139, 104469. 10.1016/j.jag.2025.104469

  • CrossRef
  • Google Scholar