LCS-Net: a lightweight architecture for efficient coastal water segmentation

Abstract

In high-resolution remote sensing imagery, near-shore water bodies typically exhibit tortuous shorelines, fragmented lakeshore coves, and superimposed disturbances such as building reflections, vegetation shadows, and mixed substrates, posing significant challenges to the fine extraction of water boundaries. Although deep learning-based semantic segmentation has substantially improved water body recognition accuracy, most existing methods focus on global context modeling while paying insufficient attention to computational efficiency on edge devices; existing lightweight models also struggle to balance high precision with low-power deployment in complex near-shore scenarios. To address this gap, this paper proposes LCS-Net, an edge-deployment-oriented architecture that establishes a system-level collaborative paradigm of “redundant feature generation—dynamic feature filtering—bypass compensation” through the recombination of existing efficient components. Following a UNet-type encoder-decoder framework, the architecture constructs a lightweight backbone by cascading standard Ghost modules with depthwise separable convolutions, embeds channel-spatial attention mechanisms to suppress spectral aliasing, and introduces bilinear-interpolation-convolution upsampling with skip connections to restore high-frequency details. Experimental results demonstrate that on a self-built unmanned aerial vehicle near-shore water dataset, LCS-Net achieves 95.92% Intersection over Union (IoU) and 97.86% F1-score with merely 2.24 million parameters and 8.24 GFLOPs, simultaneously attaining high accuracy and efficiency. It should be noted that the reported performance was obtained under specific seasonal and sensor configurations; the model’s zero-shot generalization capability under broader environmental conditions awaits subsequent validation.

1 Introduction

Coastal waters serve as vital buffer zones bridging land and sea, playing a pivotal role in disaster prevention, ecological security, and fine-scale resource management (Sun W. et al., 2023). Research indicates that these waters effectively attenuate flood peaks and buffer against coastal floods and erosion caused by storm surges and extreme rainfall, forming an essential component of integrated disaster reduction systems in coastal regions (Guo et al., 2022). Moreover, high-quality coastal water environments and well-connected water networks are fundamental to maintaining the stability of estuarine wetland habitats, biodiversity, and regional ecological security (Wang, 2019). In high-resolution remote sensing imagery, coastal waters often exhibit geometrically complex shorelines and fragmented lake branches, frequently interspersed with small land features like reefs, rafts, and aquatic vegetation, as well as complex background interference such as building reflections and vegetation shadows. These factors result in highly heterogeneous and locally blurred water boundaries.

Traditional methods relying on manual patrols and ground monitoring, while achieving high precision in localized areas, face limitations in spatial coverage and update frequency due to human resource constraints and observational conditions, making them inadequate for large-scale, continuous dynamic monitoring (Feyisa et al., 2014). On the other hand, water extraction methods based on thresholds or spectral indices (such as single-band thresholds, water indices, and their variants) demonstrate effectiveness in regions with moderate-to-low resolution and relatively simple backgrounds. However, in nearshore scenarios featuring curved shorelines, complex substrates, and frequent shadows/reflections, these methods often exhibit issues like missing boundary details, fragmented small-scale water bodies, and increased noise points (Fisher et al., 2016; Sarp and Ozcelik, 2017), failing to support refined, dynamic monitoring at watershed scales. With the widespread adoption of UAV (Unmanned Aerial Vehicle) remote sensing and edge computing devices, achieving high-precision automatic segmentation of complex nearshore water boundaries in field environments with limited computing power has become an urgent application requirement.

Deep learning-based semantic segmentation methods have provided a new technical approach for automatic identification of nearshore water bodies. Architectures such as UNet, DeepLabV3+, UNet++, and BiSeNetV2, represented by Convolutional Neural Networks (CNN), have achieved significant progress in remote sensing image object segmentation by jointly modeling multi-scale spatial information and contextual dependencies through mechanisms like encoding-decoding, atrous convolution, and feature pyramids (Huang L. et al., 2024; Lv et al., 2023; Ma et al., 2019; Wang et al., 2022). In recent years, hybrid architectures like TransUNet, which incorporate Transformer mechanisms, have further enhanced long-range dependency and global contextual modeling capabilities, improving robustness in boundary delineation and category discrimination under complex scenarios (Diakogiannis et al., 2020; Stoian et al., 2019). These methods enable finer characterization of water body morphology and shoreline details through pixel-level classification, achieving overall accuracy surpassing traditional approaches. However, most existing work primarily focuses on improving segmentation precision or enhancing global contextual modeling, with insufficient attention to computational efficiency and real-time performance on edge devices. Although a number of lightweight segmentation networks for remote sensing water bodies or general scenarios have emerged recently, reducing parameter quantities and computational overhead through backbone structure compression, pruning, and efficient operator design (Deng et al., 2025; Duan et al., 2025; Huang B. et al., 2024; Xiong et al., 2023), their designs generally remain primarily limited to backbone structure compression, offering limited support for fine representation of narrow waterways, fragmented lakes coves, and small-scale targets in nearshore environments. Some research has begun exploring the integration of lightweight encoders with attention mechanisms or multi-scale decoding modules to balance feature representation capabilities and model complexity. However, when deployed in practical applications such as UAV or embedded devices, these approaches still face challenges like high inference latency or significant accuracy degradation (Chen et al., 2024; Tong et al., 2024), making it difficult to simultaneously achieve high-precision segmentation and low-power deployment in complex nearshore scenarios. In other words, current nearshore water segmentation continues to grapple with the core issue of “difficulty in co-optimizing precision and lightweight model design.”

This core challenge manifests in nearshore scenarios through two interrelated difficulties. Firstly, spectral overlap between water and non-water objects in RGB bands occurs due to vegetation shadows, building reflections, and mixed substrates. Studies show that shallow water areas affected by vegetation and terrain shadows often exhibit similar reflectance and brightness distributions to dark land targets in visible light bands, making reliable differentiation through simple thresholds or single bands difficult (Chen and Ran, 2019). Additionally, high-reflectance reflections from buildings and vessels on still water surfaces overlap with real water boundaries in texture and edge morphology, further complicating water-land boundary identification (Thirugnanasammandamoorthi et al., 2025). In complex nearshore environments with diverse substrates and significant transparency variations, exposed tidal flats, turbid shallow waters, and suspended sediment-affected water bodies display similar tonal and textural characteristics in RGB combinations, rendering traditional water extraction methods based on color differences or spectral distance ineffective for stable segmentation (Cao et al., 2025; Liu et al., 2021). Despite existing attempts to reduce model complexity through channel compression, convolutional decoupling, and network lightweighting, these approaches often fail to adequately represent high-frequency geometric details and edge gradients, leading to frequent misclassification or omission of fine-grained targets like lake branch termini, reef edges, and aquaculture rafts (Wang Y. et al., 2021). The primary challenge for fine segmentation lies in effectively suppressing spectral overlap and enhancing complex boundary response while maintaining model lightweighting. Secondly, edge CPU devices typically lack high-performance GPU support, resulting in significant coupling between convolution operator types, network topology, and hardware execution characteristics. While some high-precision models may achieve acceptable theoretical FLOPs, their extensive use of atrous convolutions, multi-branch parallel structures, or dense connections renders them inefficient for single-threaded CPU execution. Therefore, the second challenge lies in designing a lightweight architecture compatible with edge hardware, enabling the “parameter/FLOPs advantage” to be effectively translated into real-world benefits in inference frame rate (FPS) and energy efficiency.

To address these challenges, this paper proposes LCS-Net (Lightweight Coastal Segmentation Network), a lightweight semantic segmentation model tailored for near-shore water body segmentation. Employing a UNet-like encoder-decoder architecture, the model establishes a Ghost-DSC lightweight backbone at the encoding stage by cascading Ghost modules with depthwise separable convolutions. This design achieves synergistic compression of both parameter count and computational cost through redundant feature generation coupled with spatial-channel decoupling. Regarding feature enhancement, CBAM (Convolutional Block Attention Module) is integrated into each convolutional block to dynamically recalibrate channel and spatial dimensions, thereby strengthening feature responses for water boundaries and small-scale objects while mitigating spectral aliasing. At the decoding stage, the model employs a bilinear-interpolation-convolution upsampling mechanism, complemented by cross-layer skip connections that explicitly inject shallow geometric features to compensate for high-frequency detail loss incurred by lightweight encoding. Crucially, the proposed architecture does not pursue algorithmic modifications to the Ghost modules, depthwise separable convolutions, or CBAM attention mechanisms per se; instead, it resolves the structural tension between information loss induced by lightweight encoding and subsequent decoding recovery through system-level reconfiguration of component cascading sequences and cross-layer compensation mechanisms. Through the synergistic design paradigm of “redundant generation—dynamic filtering—bypass compensation,” LCS-Net achieves substantial reduction in model complexity without compromising segmentation accuracy, thereby providing a viable deployment solution for near-shore water mapping on UAV platforms and portable edge devices.

The contributions of this work are threefold. First, to address complex near-shore aquatic environments, we construct an enhanced dataset based on high-resolution UAV imagery, systematically covering tortuous shorelines, fragmented coves, and diverse spectral interference scenarios, thereby providing high-quality sample support for evaluating the boundary-preservation and anti-confusion capabilities of lightweight models. Second, we establish a system-level compression-compensation paradigm tailored for near-shore water segmentation, distinguishing our approach from conventional lightweight UNet variants that focus solely on backbone replacement or single-module pruning. Through the cross-stage causal-chain design comprising Ghost-DSC cascaded structures, dynamic feature filtering via standard CBAM modules, and bilinear-interpolation-convolution upsampling with skip connections, the proposed architecture achieves significant model complexity reduction while maintaining high segmentation accuracy. Third, we conduct comprehensive multi-model comparisons and ablation studies on our self-built single-season Panzhihua Sun Lake (Taiyang Lake) dataset, systematically evaluate deployment performance on both edge CPU and GPU platforms, and assess transferability under limited domain shift on the public Chongming Island dataset via fine-tuning-based domain adaptation. Results demonstrate that with merely 2.24 M parameters and 8.24 GFLOPs, LCS-Net achieves 95.92% IoU, delivering 6.96 FPS on CPU and 58.42 FPS on GPU, thereby attaining a superior accuracy-efficiency trade-off compared to several mainstream methods.

2 Materials and methods

2.1 Data sources

Considering the insufficient sample size and quality of existing public nearshore water datasets, this study employed a DJI Mavic 3 RTK drone to collect remote sensing imagery of the Sun Lake nearshore zone in Hongge Town, Yanbian County, Panzhihua, Sichuan Province, in November 2024 (Figure 1). A total of 600 RGB three-band images were acquired at a spatial resolution of 0.05 m. The study area features low hilly terrain with an average lake depth of approximately 2.1 m and tortuous shorelines with multiple inlets. The nearshore region is characterized by scattered fine-scale land cover objects such as reefs, aquatic vegetation, and aquaculture rafts, creating complex geometric configurations. In this scenario, building reflections and vegetation shadows generate persistent spectral aliasing that overlaps with bottom substrate textures, rendering water-non-water discrimination difficult in localized regions and significantly complicating boundary delineation. Consequently, the Sun Lake nearshore zone is representative in terms of shoreline geometric complexity and spectral interference types, serving as a benchmark site for evaluating model boundary preservation capability and robustness to spectral interference.

FIGURE 1

2.2 Data processing

For the acquired UAV orthorectified imagery, this study employed LabelMe 5.5.0 for manual pixel-level annotation. In the binary masks, near-shore water regions were assigned a value of “1” (white) and non-water regions “0” (black) to ensure clear discrimination between water and non-water classes in the label space. During annotation, raw images exhibiting obvious exposure anomalies, motion blur, or severe occlusion were excluded to enhance overall dataset quality.

To prevent data leakage and ensure authentic evaluation of model generalization, this study adopted a “split-then-augment” data preparation protocol. First, the filtered 600 raw images and corresponding labels were divided into training, validation, and test sets at an 8:1:1 ratio. Subsequently, geometric augmentation was applied exclusively to the 480 training images, with images and labels undergoing synchronized 90°, 180°, and 270° rotations as well as vertical and horizontal flips to simulate water body morphologies under diverse capture perspectives and attitudes. The 60 images in both validation and test sets remained in their original state without any augmentation, ensuring that model performance evaluation was based on genuine unseen data.

Through this augmentation strategy, the training set was expanded to 3,360 images. Combined with the unaugmented validation and test sets, the final dataset comprised 3,480 valid samples. Data augmentation examples are illustrated in Figure 2, where (a) depicts the water body label image and (b) the corresponding original imagery.

FIGURE 2

2.3 Research methods

2.3.1 LCS-Net architecture

LCS-Net adopts a three-stage causal-chain architecture of “redundant generation—dynamic filtering—bypass compensation,” adhering to an overall five-stage encoder-decoder framework analogous to UNet. This causal chain does not represent a naive stacking of isolated modules, but rather a synergistic solution engineered to address the constraint that “compression induces information loss necessitating cross-layer structural compensation” inherent in near-shore water segmentation. Technical implementation comprises three functional components: CBAM-integrated convolutional blocks, Ghost-DSC (Ghost Module–Depthwise Separable Convolution) encoding blocks, and a bilinear-interpolation-convolution upsampling mechanism, with the overall architecture illustrated in Figure 3. Specifically, CBAM attention mechanisms are integrated within convolutional blocks to strengthen feature responses for water boundaries and small-scale objects across both channel and spatial dimensions, while suppressing background noise and spectral aliasing interference introduced during downsampling (Li et al., 2021; Song et al., 2025). Within the encoder’s feature extraction layers, Ghost modules are cascaded with depthwise separable convolutions: the former generates redundant feature maps to construct efficient representations, while the latter achieves spatial-channel decoupling to significantly reduce computational complexity (Sun Y. et al., 2023; Weng et al., 2020). At the decoding stage, a bilinear-interpolation-convolution upsampling mechanism is employed, complemented by skip connections that explicitly inject shallow geometric features into deep semantic features, thereby progressively recovering fragmented boundary details such as inlets and reefs, ultimately producing 256 × 256 pixel water segmentation outputs.

FIGURE 3

2.3.2 CBAM attention mechanism

CBAM (Convolutional Block Attention Module) is a standard channel-spatial attention mechanism proposed by Woo et al. (2018). This study does not pursue algorithmic modifications to the CBAM mechanism per se; instead, it embeds CBAM within the Ghost-DSC encoding pipeline as a “dynamic filtering” stage to suppress spectral aliasing in near-shore scenarios. In remote sensing imagery of near-shore water bodies, building reflections and vegetation shadows frequently exhibit spectral responses in RGB bands similar to those of water; such spectral aliasing significantly increases the risk of boundary misclassification. The spatial invariance inherent in conventional convolutional neural networks limits the model’s discriminative capacity in these ambiguous regions, rendering it difficult to stably highlight authentic water targets against complex backgrounds. To address this, LCS-Net integrates CBAM attention mechanisms within each convolutional block to perform recalibration of feature maps through channel and spatial attention (Liu et al., 2024). Within the system-level collaborative framework, CBAM does not function as an isolated accuracy-enhancement plugin, but rather serves as a “dynamic filtering” stage embedded in the causal chain: following redundant feature generation by front-end Ghost modules, CBAM compresses this redundant stream into effective signals highly relevant to water-body discrimination via channel-spatial recalibration, thereby providing low-noise, high-discriminability inputs for lightweight decoding at the back-end. Absent this filtering stage, the feature stream compressed by Ghost-DSC would suffer from weakened cross-channel information, rendering recovery through simple upsampling infeasible (see Exp 06–08 in Section 3.4).

The calculation process can be expressed as follows (Equations 1, 2):

In the formula: M_c and M_s represent channel and spatial attention graphs respectively, F denotes the input feature map, AvgPool and MaxPool are global average pooling and global maximum pooling respectively, MLP is a shared multi-layer perceptron that generates attention weights through nonlinear transformations, σ is the Sigmoid activation function, and f7 × 7 indicates a 7 × 7 convolution operation. While the aforementioned CBAM module performs sequential recalibration on the feature map, it inevitably introduces additional computational overhead. Therefore, LCS-Net achieves parameter compression by incorporating the Ghost-DSC structure into the encoder, which enables redundant feature generation and spatial-channel decoupling, thereby controlling overall complexity while maintaining the attention enhancement effect.

2.3.3 Ghost-DSC encoding block

The channel–spatial dual attention of CBAM enhances feature discriminability yet introduces additional convolutional kernel weights and fully-connected layer parameters, increasing the computational load on the encoder. To balance accuracy and efficiency in near-shore scenarios, LCS-Net introduces Ghost-DSC structures within core encoder layers as lightweight compensation for attention-induced overhead. It should be emphasized that both Ghost modules (Han et al., 2020) and depthwise separable convolutions (DSC) are well-established efficient computational components validated by existing literature; their theoretical foundations and implementation details do not constitute algorithmic innovations claimed by this work. The core engineering contribution of this paper lies in their cascading sequence and compression ratio configuration—Ghost modules generate redundant features via 1 × 1 convolutions at the front-end, while DSC performs computational compression through spatial–channel decoupling at the back-end, forming a causal chain of “redundant generation and decoupled compression.” Specifically, the Ghost-DSC encoding block cascades Ghost modules with depthwise separable convolution in sequence: based on the feature redundancy hypothesis, Ghost modules first generate intrinsic feature maps through primary convolutions, then produce redundant “phantom” feature maps (Han et al., 2020; Misbah et al., 2025; Shen et al., 2023; Zhang et al., 2023) via linear transformations implemented by per-channel convolutions; these are concatenated to form the complete output.

Specifically, let the number of input channels be Cin and the number of output channels be n. The parameter count of a standard k × k convolution is P_standard, while the parameter count of the Ghost module is P_ghost, which can be expressed as Equations 3, 4:where m = n/s is the number of eigen-channel and s is the compression ratio. When s = 2, the parameter compression is 1/3 of the standard convolution.

Depthwise separable convolution decouples standard convolution into two stages: depth convolution (3 × 3 per-channel) and point convolution (1 × 1 per-channel mixing). The computational complexity and theoretical compression ratio are given in Equations 5–7:

When the output channel count Cout ≥ 128, the computational load can be reduced to approximately 1/9 of the theoretical lower bound. By decoupling the computational processes of feature generation and depthwise separable convolution through the Ghost module, the Ghost−DSC architecture significantly reduces parameter count and FLOPs while maintaining optimal feature representation. In the overall architecture design, LCS−Net implements static proportional reduction of channel counts in the five-stage encoding-decoding process, compressing the standard 64–1024 channels by 25% to 48–768 channels. This channel reduction strategy is orthogonally combined with Ghost−DSC to further minimize inter-layer connection complexity while preserving multi-scale feature recognition capabilities. It should be noted that the aforementioned spatial-channel decoupling and channel reduction inevitably weaken the representation of high-frequency details during computational cost compression. Therefore, compensation for information loss is required at the decoding end through residual connections and a bilinear interpolation-convolution upsampling mechanism.

2.3.4 Bilinear interpolation-convolution upsampling and skip connection

While the Ghost-DSC encoding block achieves model compression through redundant feature generation and spatial–channel decoupling, the representation of high-frequency geometric details is inherently constrained by the nature of depthwise separable convolutions. Specifically, the channel-wise independent computation in depthwise convolution diminishes cross-channel information interaction, potentially causing edge discontinuities and region fragmentation during the upsampling reconstruction of fine-scale targets such as lake inlets and reefs in the decoding stage (Cai et al., 2025; Li et al., 2025; Zhang et al., 2024; Zhang et al., 2018). To alleviate this limitation, LCS-Net incorporates a bilinear-interpolation-convolution upsampling mechanism alongside skip connections within the decoder (Ronneberger et al., 2015).

As a substitute for conventional transposed convolution, this mechanism adopts a two-stage strategy: initially, bilinear interpolation-convolution upsampling generates a smoothed, enlarged feature map; subsequently, a 3 × 3 convolutional layer refines edge gradients and structural details. Reflection padding and LeakyReLU activation collectively ensure smooth boundary transitions with adequate contrast. The mathematical formulation of this process is given in Equations 8, 9:

In the formula, BilinearUp (F, s) denotes a bilinear interpolation operation with s as the upsampling scale factor, and f3 × 3 represents a 3 × 3 convolution layer.

Figure 4 presents a comparative analysis of traditional upsampling with post hoc convolutional refinement. Figure 4a demonstrates the original high-resolution water boundary image (exemplified at 100 × 100 pixels), exhibiting distinct inlet morphologies. Figure 4b illustrates the bilinear interpolation principle: new pixel values (marked by red dots P) are estimated through distance weighting within a 2 × 2 neighborhood. Figure 4c depicts the 3 × 3 sharpening convolution kernel employed in this study, featuring a central weight of 5 and peripheral weights of −1 to enhance high-frequency components. Figure 4d displays the result after 4 × downsampling of the original image followed by bilinear interpolation upscaling to 400 × 400: the boundary appears continuous yet overly smoothed, with blurred inlet details (yellow box). Figure 4e applies the convolution kernel from Figure 4c to the result in Figure 4d: water-edge gradients are enhanced, fine-scale inlet topologies are partially restored, accompanied by ringing effects (green box). Figure 4f presents a side-by-side comparison of local regions from Figures 4d,e, revealing differences in shoreline inflection processing. This visualization exposes the inherent trade-off between smoothing and structural fidelity in conventional decoupled upsampling, highlighting the limitations of post-filtering. Consequently, bilinear interpolation coupled with convolutional refinement is adopted as the upsampling mechanism.

FIGURE 4

On the other hand, skip connections mitigate the vanishing gradient problem in deep networks through cross-layer feature reuse while reinstating shallow geometric information. Implementation-wise, these connections are integrated within each decoding block prior to upsampling. For high-resolution feature maps output by shallow encoders, the skip pathway performs 1 × 1 pointwise convolution for channel alignment, subsequently bypassing Ghost-DSC re-encoding to inject features directly across layers. This design minimizes the risk of high-frequency detail attenuation in raw imagery.

2.3.5 Method summary

Synthesizing the aforementioned architectural components, the lightweight nature of LCS-Net is realized through a triple synergistic mechanism of “redundant generation–dynamic filtering–bypass compensation”: the front-end generates redundant features via Ghost modules to furnish adequate feature dimensionality for intermediate adaptive recalibration; the intermediate stage leverages CBAM-driven channel–spatial dual recalibration to compress this redundant stream into discriminative signals highly pertinent to near-shore water body identification; and the back-end applies cross-layer geometric injection alongside bilinear-interpolation-convolution upsampling to deliver targeted compensation for high-frequency detail loss induced by lightweight encoding, thereby achieving synergistic unification of accuracy and efficiency in edge-deployment scenarios.

3 Experimental results

3.1 Experimental configuration and evaluation criteria

All ablation experiments and control experiments in this study were conducted on the same cloud computing platform to ensure comparability of results. The hardware configuration included a single NVIDIA RTX 5090 graphics card, a 16-core Intel Xeon Platinum 8352V processor, and 90 GB of system memory. The software environment comprised the PyTorch 2.8.0 deep learning framework, Python 3.12, and CUDA 12.8, with dependencies uniformly managed through the Conda environment to ensure reproducibility of the experiments.

For training configuration, the batch size was set to 16, with 300 epochs and a patience = 50 early stopping strategy to mitigate overfitting risks. The loss function employed binary cross-entropy (BCE) with logit loss, using AdamW optimizer with a weight decay coefficient of 1e-5, while the learning rate was dynamically adjusted via cosine annealing. All comparison models maintained identical training epochs and early stopping configurations, with other hyperparameters aligned to original literature specifications.

To comprehensively evaluate the model’s segmentation performance, this study employs standard metrics in semantic segmentation: mean intersection over union (IoU), F1 score (F1), recall (R), precision (P), and overall accuracy (OA), with their formulas defined in Equations 10–14:

In the formula: TP denotes the number of pixels correctly predicted as water bodies; TN represents the number of pixels correctly predicted as non-water bodies; FP indicates the number of pixels erroneously predicted as water bodies; FN refers to the number of pixels erroneously predicted as non-water bodies.

3.2 Edge CPU deployment methodology details

This study targets edge CPU deployment scenarios by employing software-constrained single-threaded execution (PyTorch threads = 1) to evaluate performance under serial computational constraints on a multicore Intel i7-1360P processor, thereby simulating the limitations of actual edge devices. In practical near-shore water monitoring applications—specifically offline inference on UAV-embedded systems—multithreaded parallelism often introduces synchronization overhead and thread management complexity; furthermore, most embedded platforms disable hyper-threading by default or restrict active core counts to ensure thermal stability and power consumption predictability during sustained outdoor operations. Consequently, single-threaded execution represents a worst-case yet realistic deployment baseline for portable terminals, where Thermal Design Power (TDP) is strictly constrained and support for parallel computing units is unavailable.

To establish a reproducible edge performance baseline, this study conducted 1,000 independent inference stability tests on a Huawei MateBook Pro 2023 platform (Intel i7-1360P), with input dimensions fixed at 256 × 256, single-threaded execution, post-warmup measurement, and the first 10 warmup iterations excluded. The test results reveal an average inference latency of 138.84 ms (±4.89 ms), an average frame rate of 7.20 FPS, a minimum frame rate of 6.96 FPS, and a coefficient of variation (CV) of 3.52%. This empirically measured stability data (standard deviation < 5 ms across thousands of runs) validates the reliability of single-threaded testing, demonstrating that the established methodology can robustly capture the true performance boundaries of edge deployment scenarios. Constrained by the current testing environment (Huawei MateBook Pro 2023 notebook, Windows 11 operating system), this study was unable to conduct direct measurements of peak RAM usage or energy consumption per frame.

Regarding memory requirements: Based on the model architecture, storing 2.24M parameters in 32-bit floating-point format requires approximately 8.96 MB; with additional feature map buffers, the theoretical minimum memory footprint is estimated at 30–40 MB. Actual PyTorch runtime overhead may cause resident memory to significantly exceed this theoretical estimate; precise peak memory profiling requires psutil-based measurement in a future Linux environment. Regarding energy consumption: Inferred from computational density, LCS-Net requires 8.24 GFLOPs per frame compared to DeepLabv3+'s 173.39 GFLOPs—merely 4.7% of the computational load on identical platforms—enabling qualitative assessment of significant energy efficiency advantages for LCS-Net. Precise per-frame energy consumption (mJ/frame) requires measurement via Intel RAPL or dedicated power meters, which was not executed due to interface limitations of the Windows-based test device.

The aforementioned testing methodology was established on specific hardware configurations: a 13th-generation Intel Core i7-1360P processor (Huawei MateBook Pro 2023, 2.2 GHz base frequency, 5.0 GHz turbo frequency), PyTorch 2.8.0 (Intel MKL-DNN backend), single-threaded execution (threads = 1), and 256 × 256 input resolution. Performance metrics will vary with CPU microarchitecture, thermal management strategies, PyTorch backend libraries, and concurrent processes.

3.3 Comparative experiments and edge deployment validation

3.3.1 Baseline model selection strategy

The comparative experiment design in this study adheres to the principle of “precision–efficiency trade-off space coverage.” Specifically, the selected models span three characteristic design quadrants. Precision-prioritized architectures, exemplified by DeepLabV3+, pursue global context modeling through ASPP modules and deep backbones, incurring high parameter counts and computational overhead. Efficiency-prioritized architectures, represented by BiSeNetV2, employ a dual-branch structure to decouple spatial detail from semantic context in pursuit of inference speed; however, the Detail Branch’s limited effective receptive field results in noticeable accuracy degradation in complex boundary scenarios. Balanced architectures, including UNet++, DCSAUNet, and SegNet, seek equilibrium between feature reuse and computational cost through dense skip connections, attention-enhanced skip connections, and pooling index propagation mechanisms, respectively. LCS-Net targets “moderate lightweighting with high precision” under edge deployment constraints; consequently, the baseline selection aims to validate its relative position within this specific Pareto zone.

Recent proposals such as MobileNet-UNet variants, Fast-SCNN, and lightweight Transformer-CNN hybrid architectures have demonstrated notable performance in remote sensing segmentation; yet their design objective functions and deployment tiers differ fundamentally from LCS-Net (Poudel et al., 2019; Wang et al., 2022; Yuan et al., 2025). MobileNet-UNet belongs to the extreme compression category (typically < 1M parameters), with its precision–efficiency trade-off point significantly below the regime of interest in this study. Fast-SCNN’s dual-branch design targets high-resolution urban scenes, presenting receptive field adaptation issues under 256 × 256 near-shore water inputs. Even lightweight-transformed methods retain self-attention mechanisms whose memory access overhead on single-threaded CPUs fails to translate into tangible inference frame-rate advantages, and their parameter counts generally exceed 5M—representing an order-of-magnitude difference from the 2.24M edge-deployment positioning of this work. Given these architectural objective-function disparities, these models occupy distinct Pareto frontiers relative to LCS-Net; direct comparison would poorly reflect the value of the “moderate lightweighting with cross-layer geometric compensation” design paradigm pursued herein.

All comparative models underwent single-run training from scratch on identical training/validation/test splits, with unified 256 × 256 input resolution and with the same data augmentation strategies, training epochs, and early-stopping configurations as LCS-Net, ensuring comparability of precision and efficiency metrics. On the basis of unified training strategies and computational platforms, this study selects five representative semantic segmentation models—UNet++, DCSAUNet, DeepLabV3+, SegNet, and BiSeNetV2—as comparative benchmarks to systematically evaluate the relative performance of LCS-Net in near-shore water segmentation tasks.

To further validate LCS-Net’s deployability on edge devices, this study simulates extreme field offline monitoring scenarios via a PyTorch single-threaded inference environment on a 13th Gen Intel(R) Core (TM) i7-1360P processor, verifying inference speed under severely constrained computational resources with input resolution fixed at 256 × 256. Precision and efficiency metrics for all models on this platform are summarized in Table 1. Additionally, to validate upper-bound performance under abundant computational resources, supplementary desktop-level GPU (NVIDIA RTX 3060) tests were conducted: under identical 256 × 256 inputs, LCS-Net achieves 58.42 FPS, substantially exceeding the general real-time video standard of 15–30 FPS; even at 512 × 512 high resolution, it maintains 16.98 FPS, approaching the real-time threshold. These results indicate that LCS-Net’s frame-rate performance exhibits platform dependency: on edge CPUs it satisfies engineering requirements for second-level response latency, while on desktop GPUs it fully meets general real-time definitions—thereby validating the model’s cross-platform deployment flexibility.

Comparative results indicate that LCS-Net maintains competitive leadership across core accuracy metrics in near-shore water segmentation. Specifically, LCS-Net achieves 95.92% IoU, on par with the leading DeepLabV3+, while its F1-score of 97.86% markedly surpasses UNet++ at 94.70% and DCSAUNet at 96.65%. With Recall of 97.88% maintaining competitive sensitivity, and OA and Precision reaching 98.87% and 97.84%, respectively, the model demonstrates a well-calibrated balance between accuracy and robustness.

Figure 5 illustrates prediction results of the six models in a typical near-shore water area concurrently containing fragmented objects, building reflections, and vegetation shadows—conditions that place high demands on models’ ability to preserve spatial details and suppress spectral aliasing. LCS-Net’s predictions are highly consistent with ground-truth labels: even fine branches at the distal ends of coves and reef margins display continuous, crisp segmentation contours. This performance is chiefly attributed to the explicit reuse of shallow geometric features through skip connections and the amplification of edge-gradient responses by CBAM.

FIGURE 5

By contrast, although DeepLabV3+ achieves comparable IoU to LCS-Net, its ASPP module’s multi-scale feature aggregation demonstrates favorable segmentation continuity in broad water-body regions yet suffers from a certain degree of over-smoothing at the pixel level, with omission errors on some fine-scale objects—consistent with an architectural design that prioritizes global context modeling. While ASPP expands the receptive field, its explicit enhancement of fine-grained edge features remains relatively limited.

UNet++ and DCSAUNet enhance cross-layer feature reuse through dense connections; however, this may introduce redundant information, leading to varying degrees of misclassification and omission in vegetation-reflection and spectrally aliased regions shown in Figure 5. BiSeNetV2 employs a lightweight dual-branch architecture with comparatively high inference speed, yet the Detail Branch’s restricted effective receptive field compromises holistic recognition of intact water-body targets relative to the Context Branch, resulting in more pronounced false-detection artifacts in Figure 5.

Seg-Net employs a symmetric encoder–decoder architecture with pooling-index propagation, delivering relatively complete localization of principal water-body boundaries in Figure 5. However, its adaptability to locally fragmented shorelines and fine-textured regions is limited, where boundary jitter and blurring readily emerge. Synthesizing these observations, divergent architectures exhibit differentiated trade-offs along the accuracy–lightweighting axis: benchmarked against LCS-Net, DeepLabV3+ maintains comparable IoU through substantially elevated parameter and computational budgets; BiSeNetV2 sacrifices 5.00 percentage points of IoU in exchange for higher inference speed; while intermediate solutions such as UNet++ strike varying balances between feature reuse and computational cost. These architectural disparities are ultimately concentrated in execution efficiency within edge deployment scenarios.

Inference speed on edge CPUs is fundamentally a matter of architectural design matching hardware execution characteristics, rather than a simple mapping from FLOPs or parameter counts. LCS-Net achieves 6.96 FPS because the cascaded Ghost modules and depthwise separable convolutions align with the serial nature of CPU execution: Ghost modules are dominated by 1 × 1 pointwise convolutions—a regular computational pattern readily vectorized by CPU SIMD instruction sets—while DSC decomposes standard convolutions into depthwise and pointwise stages, reducing cross-channel computational overhead. Their synergy compresses computational density to 8.24 GFLOPs, enabling 6.96 FPS on the i7-1360P. In contrast, DeepLabV3+’s ASPP module employs atrous convolutions with varying dilation rates; under single-threaded execution, kernels with different dilations access feature maps at irregular intervals. Although its IoU matches LCS-Net, the 173.39 GFLOPs constrain execution efficiency on the i7-1360P, yielding only 2.70 FPS—merely 39% of LCS-Net. This demonstrates that high-compute architectures misaligned with the serial characteristics of edge hardware fail to translate theoretical advantages into practical performance. BiSeNetV2 compresses computation to 11.18 GFLOPs and achieves 6.60 FPS, close to LCS-Net; however, its dual-branch design cannot be parallelized in a single thread, instead incurring additional computational overhead from inter-branch synchronization. More critically, the receptive field sacrificed for speed causes its IoU to drop by 5.00 percentage points. This illustrates that on single-threaded CPUs, such dual-branch architectures improve frame rates through computational compression yet sacrifice segmentation accuracy due to design compromises, failing to achieve synergistic accuracy–efficiency optimization. UNet++’s dense skip connections enhance feature reuse but introduce complex computational dependencies during serial execution; its high computational density of 139.61 GFLOPs limits FPS to 5.93. Although DCSAUNet’s attention-enhanced skip connections reduce computation to 27.66 GFLOPs, the attention mechanism likewise introduces cross-layer synchronization overhead under single-threaded execution, yielding only 5.82 FPS—failing to open a performance gap commensurate with its computational advantage. SegNet, constrained by its symmetric encoder–decoder structure and pooling-index propagation mechanism, requires frequent memory-access operations in single-threaded mode; the 160.52 GFLOPs computational burden further degrades CPU execution efficiency, resulting in merely 3.46 FPS—only marginally above DeepLabV3+’s 2.70. Furthermore, on a notebook platform equipped with an RTX 3060, LCS-Net reaches 58.42 FPS, with per-image latency reduced to 17.12 ms, validating the architecture’s adaptation potential and deployment flexibility on GPU platforms.

These analyses are intuitively corroborated by Figures 6a,b: LCS-Net sits in the favorable upper-left region of the Pareto Frontier across both the accuracy–computation and parameter–FPS dimensions, showcasing its comprehensive superiority in edge deployment.

FIGURE 6

Collectively, mainstream models generally fail to achieve a satisfactory balance between accuracy and lightweight edge deployment. By contrast, through the synergistic design of a Ghost-DSC lightweight backbone, attention-based feature enhancement, and cross-layer geometric compensation, LCS-Net achieves 6.96 FPS while maintaining 95.92% IoU, thereby attaining effective synergy between accuracy and efficiency in edge scenarios.

3.4 Ablation experiment

The structural contradiction in nearshore water segmentation—that accuracy and efficiency are inherently difficult to reconcile—renders component-level trade-offs a pivotal concern in architectural design. To systematically dissect the synergistic mechanisms among attention enhancement, lightweight compression, and cross-layer compensation within LCS-Net, we designed nine ablation experiments. The baseline adopts a standard five-stage encoder–decoder architecture, whereas experimental groups incrementally integrate CBAM, Ghost-DSC, bilinear interpolation-convolution upsampling, and channel reduction strategies upon this foundation, progressively assessing the marginal contributions of each module to boundary preservation and parameter compression. All experiments were conducted under identical data splits and training configurations to quantify the practical impact of differing component combinations on overall performance. Additionally, each of the nine ablation configurations underwent five independent training–testing cycles; Table 2 summarizes the mean values of each metric across these five runs.

To establish a performance evaluation benchmark, we first assessed the unoptimized standard five-stage encoder–decoder architecture (Exp01). This model employs conventional convolutional blocks with a 64–1024 channel configuration, reaching 35.57M parameters and 197.91G FLOPs. In the nearshore water segmentation task, it achieved an IoU of 95.84% and an F1-score of 97.82%, indicating that the baseline architecture possesses preliminary discriminative capability. However, the computational resource demands of this model constrain its edge-deployment potential, and this accuracy level can be further improved through subsequent component optimization. Consequently, the structural contradiction between the baseline model’s computational cost and its accuracy potential provides the rationale for introducing lightweight and feature-enhancement strategies.

Building upon this baseline, we first evaluated conventional feature-enhancement pathways. Results from Exp02 to Exp04 reveal a clear commonality: individually, CBAM, bilinear interpolation-convolution upsampling, and their combination elevated IoU to 96.18%, 95.99%, and 96.24%, respectively, yet all incurred synchronous increases in parameter count and FLOPs. These findings indicate that while attention mechanisms and edge refinement can effectively improve model recognition accuracy, they inevitably aggravate computational burdens, perpetuating the dilemma where accuracy and efficiency remain difficult to reconcile in conventional architectures.

However, the anomalous behavior exhibited by Exp05 contradicts the conventional expectation of a positive correlation between accuracy and computational cost. Following a 25% channel reduction from 64 to 1024 to 48–768, parameters decreased to 20.01M and FLOPs dropped to 111.44G, yet IoU unexpectedly rose to 96.71%—the highest value observed across all ablation experiments. This counterintuitive phenomenon can be explained through representation learning theory: the standard UNet’s 64–1024 channel configuration exhibits structural over-parameterization in shallow stages, generating numerous “inert” feature dimensions that contribute minimally to water–non-water discrimination. These redundant dimensions not only inflate computational burdens but also degrade the condition number of the gradient covariance matrix, thereby interfering with the backpropagation of informative features. Upon compression to 48–768 channels, the network is compelled to focus on high-information discriminative features, which paradoxically improves the gradient signal-to-noise ratio and engenders an implicit regularization effect. This confirms that, for this specific nearshore water segmentation task, model capacity and accuracy are not monotonically positively correlated; instead, a regularization-induced accuracy peak exists.

The optimization observed in Exp05 was predicated upon conventional convolutional architectures; its accuracy gain stemmed from implicit regularization induced by eliminating redundant channels, yet this optimization proves valid only within computationally intensive architectures. When lightweight designs are subsequently introduced, multi-module collaborative compensation becomes imperative. It must be emphasized that Exp05’s accuracy gain rested upon the contextual interaction capabilities inherent to traditional dense convolutions—its essence lies in pruning redundant dimensions. By contrast, the compression introduced by the Ghost-DSC architecture severs cross-channel information interactions (specifically during the depthwise convolution stage) atop this foundation, rendering the underlying mechanisms of feature representation fundamentally distinct between the two.

Data from Exp06 through Exp08 reveal that Ghost-DSC compression persistently reduced IoU by 1.37 to 0.93 percentage points relative to Exp05. Even with the incorporation of CBAM attention, the isolated precision recovery it delivered remained marginal, with cumulative gains failing to reach 0.50 percentage points. This demonstrates that the information degradation triggered by lightweight compression cannot be reversed through singular enhancement mechanisms.

Exp06 introduced Ghost modules atop the Exp05 foundation while retaining transposed convolutions at the decoding end, serving to validate the outcome of an isolated design strategy. Parameters plummeted from 20.01M to 5.03M, FLOPs contracted to 14.29G, yet IoU regressed to 95.34%—a 1.37-percentage-point decline relative to Exp05. The underlying cause lies in the Ghost module’s phantom-feature generation mechanism, which relies upon per-channel linear transformations lacking cross-channel information interaction, thereby diluting high-frequency edge components during the encoding phase; conventional transposed convolutions prove incapable of compensating for such losses. This result substantiates that Ghost modules and channel-aggregation upsampling must be engineered as an integrated unit: the former executes per-channel compression at the encoding stage, while the latter performs cross-channel aggregation at the decoding stage, constituting an encoder–decoder causal chain within LCS-Net.

Exp07 further incorporated DSC modules into both the encoder and skip connections, reducing parameters from 5.03M to 3.27M in exchange for a mere 0.44-percentage-point IoU recovery to 95.78%—still falling markedly below the Exp05 benchmark of 96.71%. This marginal accuracy gain indicates that while depthwise separable convolutions compress FLOPs to 11.52G through spatial–channel decoupling, their per-channel independence simultaneously undermines discriminative power over target regions, failing to fully resolve the information attenuation inherent to lightweight designs. Exp08 subsequently embedded CBAM attention, nudging parameters to 3.39M while IoU remained flat at 95.78% with Recall essentially unchanged at 97.81%. These two sets of data collectively reveal that the Ghost + DSC + CBAM combination, despite reducing computational cost by 89.7% relative to Exp05, delivered cumulative IoU gains of merely 0.44 percentage points—far from the Exp05 baseline.

Mechanistically, DSC and bilinear-interpolation-convolution upsampling are architecturally coupled and resist independent ablation. In Exp06–Exp08, transposed convolutions persisted at the upsampling stage following DSC introduction, lacking explicit geometric compensation mechanisms; consequently, lightweight-induced information attenuation could not be reversed through singular enhancement means. This indicates that when explicit geometric compensation is absent at the decoding stage, both attention mechanisms and spatial-fidelity modules fail to compensate for high-frequency losses, rendering accuracy and lightweight design difficult to reconcile.

Exp09 ultimately achieved a balance between accuracy and efficiency through the full synergy of Ghost-DSC compression, CBAM dynamic screening, channel reduction regularization, and bilinear interpolation-convolution upsampling. With 2.24M parameters and 8.24G FLOPs, it improved IoU by 0.08 percentage points relative to the baseline model, while reducing parameters by 93.71% and floating-point computations by 95.84%, thereby realizing the accuracy–efficiency trade-off.

Notably, the 96.71% IoU of Exp05 represents the theoretical accuracy upper bound, whereas the 95.92% IoU of Exp09 constitutes the engineering-deployable configuration. The 0.79-percentage-point discrepancy between these two values precisely quantifies the inherent representational loss attributable to Ghost-DSC lightweight encoding—specifically, the weakening of high-frequency edge details resulting from cross-channel information decoupling. While this loss is partially mitigated through CBAM attention-based screening and bilinear interpolation-convolution upsampling geometric compensation, it cannot be fully neutralized.

To intuitively illustrate the substantive nature of this 0.79% accuracy discrepancy, Figure 7 presents a comparison between Exp-05 (theoretical accuracy upper bound, 96.71% IoU, 20.01M parameters) and Exp-09 (engineering deployment configuration, 95.92% IoU, 2.24M parameters) across typical scenarios. As shown in Figure 7a, Exp-05 remains almost unaffected by canopy shadows, accurately delineating shoreline contours through the shadowed areas; in Figure 7b, Exp-05 demonstrates significantly superior recognition of shadows cast by artificial structures (pedestrian bridges) compared to Exp-09. Figure 7c reveals that both configurations achieve satisfactory identification of aquatic vegetation, though Exp-05 maintains an advantage in suppressing building shadows; Figure 7d indicates comparable performance between the two in spectrally confused regions.

FIGURE 7

The above comparison confirms that the 96.71% IoU of Exp-05 indeed represents superior segmentation quality, particularly in terms of shadow suppression and building contour preservation. However, this 0.79-percentage-point accuracy advantage is attained at the cost of 20.01M parameters (8.9 × those of Exp-09) and 111.44G FLOPs (13.5 × those of Exp-09). Exp-09 does not pursue absolute accuracy surpassing that of Exp-05; rather, under the stringent constraints of edge deployment, it trades a 0.79% accuracy compromise for 93.7% parameter compression relative to the baseline model and inference speeds of 6.96 FPS on CPU and 58.42 FPS on GPU. This trade-off reveals the engineering reality of nearshore monitoring: in the context of edge deployment, the 95.92% achieved with fewer parameters holds greater scenario-adaptive value than the 96.71% attained with larger parameter counts.

Unlike existing lightweight UNet variants (e.g., MobileNet-UNet, Ghost-UNet) that focus primarily on backbone replacement or single-module pruning, the differentiation of LCS-Net lies in its cross-stage causal-chain design: the Ghost-DSC compression at the encoding stage inevitably induces cross-channel information loss, which cannot be reversed through simple attention stacking (as evidenced by Exp06–Exp08), but rather necessitates structural compensation via bilinear-interpolation-convolution upsampling at the decoding stage. This “compression-compensation” cross-layer synergy, rather than module replacement, constitutes the fundamental distinction from conventional lightweight UNets.

Of course, even the 95.92% accuracy does not imply perfect performance across all scenarios. Exp-09 still exhibits specific failure modes under the physical constraints of single RGB modality input, which will be further discussed in the limitations analysis below.

3.4.1 Limitations

Despite achieving 95.92% IoU on the overall test set, LCS-Net still exhibits identifiable failure modes under specific challenging scenarios. These failure cases not only delineate the method’s boundaries of applicability but also provide an empirical window into the inherent trade-offs of lightweight design. As shown in Figure 8, four typical error patterns can be attributed to two distinct levels.

FIGURE 8

3.4.1.1 First-level attribution

Physical limitations of RGB unimodal input. All errors shown in Figures 8a–d share a common underlying premise—the model relies solely on visible RGB bands as input. This leads to three types of spectral confusion: (1) hard shadows cast by artificial structures (e.g., pedestrian bridges): cause water spectral features to be completely overwhelmed by geometric occlusion (Figure 8a); (2) turbid water bodies and riparian vegetation exhibit highly similar reflectance characteristics in RGB space (Figure 8b); and (3) radiometric attenuation induced by dense canopy shadows blurs the water-land boundary gradients (Figure 8d). These confusions constitute inherent physical limitations of the RGB modality, particularly pronounced under visible-light unimodal conditions. Consequently, the errors illustrated in Figure 8 do not represent failures unique to LCS-Net, but rather delineate the applicable boundaries of semantic segmentation models based on RGB input (Lee et al., 2025; Wieland et al., 2019).

3.4.1.2 Second-level attribution

Additional sensitivity introduced by the Ghost-DSC lightweight architecture. Key data from the ablation experiments provide structural-level explanations for this phenomenon: following the introduction of Ghost-DSC encoding blocks (Exp06--Exp08), IoU exhibits an irreversible degradation of 0.93%–1.37% even when the input modality remains identical. This degradation is particularly pronounced in the water–reef–shoal hybrid region illustrated in Figure 8c. The phantom feature generation mechanism of Ghost modules relies upon per-channel linear transformations, while the depthwise convolution stage further severs cross-channel information interaction. For low-contrast boundaries distinguishable only through cross-channel higher-order statistical differences (e.g., clear shallow water versus bright reefs), this channel-independent computational characteristic directly undermines the model’s discriminative power. Consequently, the fragmented segmentation observed in this region by LCS-Net (Figure 8c) stems not merely from RGB limitations, but represents the inherent cost of Ghost-DSC lightweight compression.

Notably, the 0.79-percentage-point IoU discrepancy between Exp05 (uncompressed, 96.71% IoU) and Exp09 (full LCS-Net, 95.92% IoU) represents the statistical quantification of the challenging scenarios illustrated in Figure 7. Put differently, LCS-Net trades less than 5% of the computational cost and 6% of the parameters (compared to baseline Exp01) for deployment feasibility on edge devices. The bulk of this precision cost is localized to the low-contrast regions described above, where cross-channel information has been weakened.

The analysis of these limitations not only delineates the applicable boundaries of LCS-Net—demonstrating robust performance in clear to moderately turbid water bodies without significant shadow occlusion, while requiring cautious deployment in spectrally confused or low-contrast boundary regions—but also clarifies that the architecture’s “accuracy-efficiency” trade-off entails no cost-free gains; rather, it is realized at the expense of sacrificing partial cross-channel high-order representational capacity. This limitation carries critical implications for practical deployment: in edge scenarios with stringent computational constraints, LCS-Net offers a precisely calibrated, high-efficiency solution under controlled precision loss, rather than a universal solution that outperforms high-compute models across all dimensions.

3.5 Cross-scenario transfer validation via fine-tuning

Fine-tuning-based domain adaptation capability constitutes a critical dimension for evaluating the practical engineering value of the model. To validate this capability for LCS-Net in heterogeneous environments, we select the publicly available Chongming Island near-shore water dataset from the Paddle AI Studio platform as the validation benchmark, comprising a total of 1,237 images.

Compared with the Panzhihua Sun Lake dataset, the Chongming Island data exhibits systematic disparities in imaging conditions, spectral characteristics of ground objects, and shoreline morphology. Chongming Island represents an estuarine alluvial landform subject to tidal dynamics, where the spatial heterogeneity of water turbidity is significantly intensified; furthermore, the presence of numerous small floating rafts and complex background interference in the imagery exacerbates spectral confusion. This substantial domain gap provides a challenging test scenario for evaluating the model’s transfer robustness.

The transfer experiment adopts a fine-tuning strategy: initially, the optimal weights trained on the Sun Lake dataset are loaded, followed by continued training for 30 epochs on the Chongming Island training set. During this process, the initial learning rate is reduced to 1e-5, BatchNorm layer statistics are frozen while maintaining the γ and β parameters as trainable, and remaining configurations—including batch size = 16, AdamW optimizer, weight decay of 1e-5, and CosineAnnealingLR scheduling—remain consistent with the source domain training. DeepLabV3+ and BiSeNetV2 are selected as comparative models, representing complex architectures with comparable accuracy levels and efficient models with similar lightweight characteristics, respectively.

Table 3 presents the performance metrics of the three models after transfer learning. LCS-Net achieves an IoU of 93.27% in the Chongming Island scenario, decreasing by only 2.65 percentage points compared to the source domain (Sun Lake). The combination of an F1-score of 96.34%, Recall of 99.12%, and Precision of 93.96% indicates that the model maintains high recall for water targets, albeit with some degree of false positive detections in background regions.

By contrast, DeepLabv3+ exhibits an IoU degradation to 90.75%, representing the largest decline of 5.32 percentage points among the three models. Its Recall concurrently drops to 92.24%, while Precision increases to 98.08%, indicating a more conservative prediction strategy adopted under the pronounced mixed-background interference characteristic of the Chongming Island scenario. BiSeNetV2 records an IoU of 88.19%, the lowest among the three; although its Recall remains relatively high at 96.98%, its Precision merely reaches 90.68%, yielding a decreased F1-score of 93.72%. This corroborates the insufficient recognition stability of its lightweight dual-branch architecture under cross-domain transfer conditions.

In summary, through the cascaded integration of Ghost-DSC compression and CBAM-based dynamic feature selection, LCS-Net achieves the most favorable cross-domain accuracy retention under the constraint of merely 2.24M parameters, exhibiting superior stability under transfer learning compared to architectures of equivalent accuracy and lightweight classes.

Figure 9 presents a comparative visualization of prediction results for typical Chongming Island regions. This area simultaneously encompasses turbid water bodies, dense floating rafts, and complex shorelines, imposing more stringent demands on the model’s spatial detail preservation and anti-interference capabilities. LCS-Net demonstrates superior performance in floating raft recognition and the maintenance of complex shoreline geometric morphology; however, it still exhibits minor misclassifications in local regions sharing spectral characteristics similar to those of Sun Lake, and local vegetation shadows further induce segmentation errors at individual water boundaries.

FIGURE 9

DeepLabv3+ demonstrates relatively complete overall boundary localization, yet lacks sufficient fine-grained delineation for highly curved or fragmented shorelines, with some rafts exhibiting fragmentation or adhesion artifacts. The shoreline contours output by BiSeNetV2 are generally consistent with the ground truth; however, both the response intensity to diminutive targets and the continuity of fine details are inferior to those of LCS-Net. Collectively, Figure 9 demonstrates that LCS-Net achieves a relative equilibrium between cross-domain accuracy and detail preservation under identical fine-tuning epochs.

To contextualize LCS-Net’s cross-domain performance within published literature, comparisons are made against existing reported results on Chongming Island and comparable estuarine datasets. ISUNet (2024) employs an improved UNet++ hierarchical skip connection architecture, trained and tested locally on Chongming Island near-shore water bodies, achieving 94.79% IoU, 97.23% F1, 98.36% OA, and 98.37% Precision with approximately 20M parameters; the Recall metric is not reported in the original source (Miao et al., 2024). DAU-Net (2023) targets the turbid intertidal water environment of the Yangtze River estuary (sharing the same estuarine alluvial geomorphology as Chongming Island), achieving 95.20% IoU, 97.50% F1, 97.60% Recall, and 97.40% Precision in local scenarios through embedded coordinate self-attention mechanisms and a Hybrid loss function, with approximately 15M parameters; OA is not explicitly reported in the original source (Chen et al., 2023). These metrics are compared in Table 4.

As shown in Table 4, under these differing experimental settings, the IoU gap between LCS-Net and ISUNet is constrained within 1.52 percentage points, while the gap relative to DAU-Net stands at 1.93 percentage points; its F1-score registers approximately 0.89–1.16 percentage points below the two aforementioned studies, attributable to representation degradation associated with the lightweight architecture and cross-domain transfer. Notably, LCS-Net attains a Recall of 99.12%, exceeding DAU-Net’s 97.60%, indicating that the cross-domain transfer strategy biases the model toward expanding the predicted water body extent to minimize false negatives (omissions)—albeit at the cost of reduced Precision (93.96%) compared to ISUNet (98.37%) and DAU-Net (97.40%), consequently elevating the false alarm rate. Given that LCS-Net’s parameter count (2.24M) constitutes merely 11.2% of ISUNet’s and 14.9% of DAU-Net’s, coupled with the inherent challenges of cross-domain distribution shift, these results demonstrate that LCS-Net maintains generalization potential comparable to reported literature methods while preserving lightweight edge-deployment capabilities; its Recall-OA trade-off profile offers substantial engineering value for monitoring applications prioritizing high water-body detection rates.

3.5.1 Dataset bias and deployment risks

The primary validation of this study is based on the Panzhihua Sun Lake dataset, the acquisition of which is subject to structural constraints of single-season, single-sensor, and single-water-quality state: data collection occurred in November 2024 (late autumn in the Northern Hemisphere), within a low-mountain hilly geomorphology characterized by an average lake depth of approximately 2.1 m, tortuous shoreline morphology, and a fragmented distribution of coves; the near-shore zone is scattered with fine-scale objects including reefs, aquatic vegetation, and artificial floating rafts. During this period, the water body exhibited a distinct green coloration with medium transparency (rather than clear visibility to the bottom), while aquatic vegetation remained in the terminal growth phase, yielding spectral features dominated by high-greenness reflectance. The sensor was fixed as the DJI Mavic 3 RTK (RGB three-band, 0.05 m spatial resolution), excluding multispectral or SAR modalities. This configuration introduces systematic spectral–geometric bias: throughout the training process, the model predominantly learned feature combinations of “green water body–fragmented shoreline morphology–specific illumination geometry,” forming a representational preference for high-greenness spectral responses and complex shoreline textures; however, this preference simultaneously constitutes a representational blind spot—the model lacks experience with clear transparent water bodies and geometric morphologies covered by winter-withered vegetation, potentially leading to discrimination errors when encountering scenarios with such distribution shifts.

The current dataset fails to cover critical regions in the season-illumination-turbidity covariate space, constituting threefold limitations. Seasonal phenological blind spots: The November acquisition missed peak-growing-season floating-leaf vegetation such as water hyacinth (Eichhornia crassipes) and duckweed (Lemna spp.), leaving the model’s discrimination capability for vegetation-water mixed pixels unverified; differences in inundation extent between dry and wet seasons may induce boundary drift under scenarios of seasonal water expansion. Illumination geometry bias: The combination of fixed vertical aerial photography and specific solar elevation angles imparts directional dependency to the learned illumination-shadow relationships; when encountering complex illumination conditions or oblique perspective distortion, the geometric morphology of building reflections undergoes significant alterations, potentially triggering systematic false detections. Turbidity gradient gaps: Sun Lake and Chongming Island represent the two poles of eutrophic green water and estuarine alluvial turbid water, respectively, with validation lacking for transition zones of intermediate trophic states; theoretical blind spots persist regarding the model’s response thresholds to spectral confusion between suspended sediments and water bodies across different turbidity intervals.

All cross-domain validations in this study were conducted via supervised fine-tuning. Zero-shot testing was deliberately omitted due to spectral non-overlap across domains. Imposing zero-shot evaluation would inevitably yield substantially depressed IoU scores; however, such results would be diagnostically uninformative, as they conflate intrinsic architectural deficiencies with failures stemming from domain gaps exceeding the model’s representational capacity. Given these inherent dataset biases, LCS-Net harbors systemic failure risks in operational deployment: inference without domain adaptation on winter ice-covered waters, high-turbidity flood zones (>100 mg/L suspended solids), nocturnal thermal infrared imagery, or data from heterogeneous sensors may trigger precipitous accuracy degradation. These deployment risks can be mitigated through target-domain fine-tuning with limited samples (<100 images); future enhancements to generalization may integrate unsupervised pre-training on unlabeled data and multi-source domain training strategies.

4 Discussion

The LCS-Net proposed in this study aims to alleviate the bottleneck where accuracy and lightweight deployment are difficult to reconcile in nearshore water segmentation. Experimental results demonstrate that by integrating channel and spatial attention mechanisms at the encoding stage, cascading Ghost modules with depthwise separable convolutions to construct lightweight encoding blocks, and employing bilinear interpolation-convolution upsampling combined with skip connections at the decoding stage, the constructed model achieves 95.92% IoU and 97.86% F1-score on our self-constructed dataset, while compressing parameters and computational costs to 2.24M and 8.24 GFLOPs, respectively, attaining 6.96 FPS inference speed on edge CPU devices. These results validate the efficacy of the “redundancy generation–dynamic filtering–bypass compensation” synergistic design paradigm, establishing the feasibility of maintaining high-precision segmentation while substantially enhancing computational efficiency.

To further evaluate the upper bound of real-time monitoring capability, this study supplemented tests on a desktop-grade GPU (NVIDIA RTX 3060): at the identical 256 × 256 input resolution, LCS-Net achieved 58.42 FPS, substantially exceeding the universal real-time video criterion of 15–30 FPS; even at 512 × 512 high resolution, it sustained 16.98 FPS, approaching the real-time threshold. These findings demonstrate that LCS-Net’s frame-rate performance is platform-dependent: delivering second-level responsiveness suitable for engineering applications on edge CPUs, while fully satisfying general real-time criteria on desktop GPUs—thereby validating the model’s cross-platform deployment flexibility.

Compared with existing mainstream models, LCS-Net demonstrates superior comprehensive performance in the accuracy-efficiency tradeoff. While models such as DeepLabv3+ possess strong feature extraction and context modeling capabilities, their substantial parameter counts and computational overhead significantly constrain real-time deployment in edge environments (Chen et al., 2018); conversely, lightweight models like BiSeNetV2, in their pursuit of inference speed, exhibit varying degrees of precision degradation for complex boundaries and fine-grained targets (Yu et al., 2021). LCS-Net achieves targeted performance equilibrium through synergistic component configuration tailored to nearshore scenarios: on one hand, the Ghost-DSC cascade substantially compresses computational cost while preserving discriminative capability; on the other hand, CBAM amplifies feature responses at edges and fine-grained targets, while geometric information re-injection at the decoding stage further mitigates high-frequency detail loss induced by lightweight encoding (Han et al., 2020; Woo et al., 2018). Moreover, the 0.79% IoU decrease from Exp09 relative to Exp05, alongside the maximum 1.37 percentage point precision loss observed in Exp06-08 upon Ghost-DSC introduction, both conform to the classical principle of representation capacity degradation attendant to model compression. Experimental evidence indicates that LCS-Net, via architectural-level synergistic design, compresses the cost of this tradeoff within manageable bounds, thereby furnishing a principled lightweight paradigm for nearshore water segmentation in this specific scenario.

Ablation experiments reveal the inherent synergistic relationships among components at the modular level. Merely applying channel reduction to the baseline model yields substantial accuracy gains, indicating pronounced structural redundancy in the standard encoder’s 64-to-1024 channel configuration (Wang Z. et al., 2021); moderate channel compression thus exerts an implicit regularization effect. However, when more aggressive lightweight components such as Ghost-DSC are superimposed upon this foundation, mere structural compression incurs representation capacity degradation, manifested as IoU decline. Under these conditions, recovery—or even surpassing—of baseline accuracy necessitates the concerted action of CBAM’s dynamic feature filtering and explicit geometric compensation at the decoding stage (Badrinarayanan et al., 2017). This suggests that in lightweight network design, information compression in the encoder and information recovery in the decoder should be treated as an integrated optimization problem rather than isolated local improvements.

Cross-scenario transfer experiments demonstrate that, upon loading pre-trained weights and executing target-domain fine-tuning, LCS-Net can adapt to estuarine environments with spectral discrepancies, exhibiting limited domain adaptation capability. On the Chongming Island dataset—characterized by pronounced differences in spectral signatures and land-cover environments—the model sustains 93.27% IoU after continued training, with its performance degradation magnitude smaller than that of comparative models. This outcome owes partly to the lightweight structure’s mitigation of overfitting risks during fine-tuning, and partly to the feature adaptive selection capability conferred by the attention mechanism (Al-Najjar et al., 2019; Guo et al., 2023; Rostami et al., 2019; Tong et al., 2020). Meanwhile, transfer experiments expose current limitations: under RGB imagery, the model’s discriminative capacity remains insufficient for spectrally confused regions encompassing water-shadow interfaces and dark terrestrial features, manifesting as Precision decline in cross-domain testing. This deficiency stems primarily from insufficient spectral diversity in training data and partial loss of high-level semantic information during lightweight encoding, suggesting the necessity of incorporating multispectral/multimodal information or integrating domain adaptation and domain generalization methods to further enhance robustness in complex scenarios.

Collectively, LCS-Net furnishes an engineering-feasible solution for lightweight fine-grained segmentation of nearshore water bodies, achieving a balance between accuracy and efficiency on edge devices through structural synergy. Nevertheless, this study exhibits several limitations: first, the model demonstrates certain dependency on specific data distributions, lacking systematic validation across larger-scale, multi-type water body scenarios; second, the current framework does not explicitly incorporate temporal information, constraining its capacity to characterize water body processes with pronounced dynamic characteristics; third, hardware deployment has focused primarily on general-purpose CPU/GPU platforms, with deep synergy with dedicated edge computing chips remaining to be explored. Future research may advance along the directions of multi-modal data fusion, temporal sequence modeling, domain adaptation enhancement, and coordinated optimization with heterogeneous edge computing resources (Fayaz et al., 2024; Hong et al., 2021), to further facilitate the practical deployment of lightweight models in production and regulatory scenarios.

5 Conclusion

Addressing the critical challenge where accuracy and efficiency are difficult to reconcile in fine-grained nearshore water monitoring under edge computing constraints, this study constructs LCS-Net, a lightweight synergistic architecture integrating existing efficient components, and delivers systematic validation from both architectural design and experimental perspectives. Through system-level synergistic configuration of components tailored to edge scenarios, LCS-Net demonstrates the engineering feasibility of maintaining 95.92% IoU while achieving 2.24M parameter compression via the “redundancy generation–dynamic filtering–bypass compensation” causal-chain architecture, establishing the engineering principle that encoding compression and decoding recovery must be optimized as an integrated causal chain in lightweight design. The theoretical contributions of this architecture manifest in two dimensions: first, revealing the redundancy inherent in standard encoder channel configurations; second, establishing the imperative of treating encoding compression and decoding recovery as holistically optimized causal chains. Experimental results indicate that LCS-Net achieves 6.96 FPS on pure-CPU edge devices, evidencing its relative efficiency advantage in compute-constrained scenarios, and attains 58.42 FPS on desktop GPUs (RTX 3060), exceeding the universal real-time video standard of 15–30 FPS. This cross-platform adaptability validates the flexibility of the “redundancy generation–dynamic filtering–bypass compensation” architecture across heterogeneous compute tiers, furnishing a viable solution spanning the full application spectrum from field offline monitoring (CPU) to desktop real-time processing (GPU).

Empirical results demonstrate that, on the self-constructed nearshore water dataset, LCS-Net achieves a parameter reduction of approximately 94.3% relative to DeepLabv3+ while maintaining statistically comparable IoU performance. In the Chongming Island cross-scenario transfer experiments, LCS-Net exhibits performance degradation of merely 2.65 percentage points—significantly outperforming comparative models—thereby validating the fine-tuning adaptability of lightweight architectures under limited domain shift conditions. Edge deployment experiments further demonstrate that cascading Ghost modules with depthwise separable convolutions effectively compresses computational cost to 8.24 GFLOPs, achieving the highest inference frame rate among all comparative models on the i7-1360P platform, thus furnishing a readily deployable algorithmic prototype for offline monitoring scenarios involving UAVs and portable terminals.

In summary, this study establishes a “compression–compensation” technical framework for lightweight semantic segmentation tailored to edge scenarios, furnishing an extensible foundation for subsequent research on multi-spectral and multi-modal information fusion, temporal process characterization, and coordinated optimization with dedicated edge chips. Future endeavors may, while preserving the lightweight characteristics of the model, further incorporate strategies such as self-supervised pre-training and cross-domain contrastive learning to enhance the generalization capability and practical value of LCS-Net in large-scale, multi-type water body monitoring tasks.

References

• 1

  Al-NajjarH. A. H.KalantarB.PradhanB.SaeidiV.HalinA. A.UedaN.et al (2019). Land cover classification from fused DSM and UAV images using convolutional neural networks. Remote Sens.11, 1461. 10.3390/rs11121461

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  BadrinarayananV.KendallA.CipollaR. (2017). SegNet: a deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell.39, 2481–2495. 10.1109/TPAMI.2016.2644615

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  CaiJ.TaoL.LiY. (2025). CM-UNet++: a multi-level information optimized network for urban water body extraction from high-resolution remote sensing imagery. Remote Sens.17, 980. 10.3390/rs17060980

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  CaoQ.ZhaoB.LiZ.ZhangF.WuY. (2025). A lightweight network with embedded soft constraints on approximate spectral features for real-time water body segmentation in remote sensing images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.18, 10608–10624. 10.1109/JSTARS.2025.3552783

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  ChenJ.RanX. (2019). Deep learning with edge computing: a review. Proc. IEEE107, 1655–1674. 10.1109/JPROC.2019.2921977

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  ChenL.-C.ZhuY.PapandreouG.SchroffF.AdamH. (2018). “Encoder-decoder with atrous separable convolution for semantic image segmentation,” in Proceedings of the 15th European Conference on Computer Vision—Eccv 2018, Munich, Germany, September 8–14, 2018, 833–851. 10.1007/978-3-030-01234-2_49

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  ChenC.ZhangC.TianB.WuW.ZhouY. (2023). Mapping intertidal topographic changes in a highly turbid estuary using dense Sentinel-2 time series with deep learning. ISPRS J. Photogramm. Remote Sens.205, 1–16. 10.1016/j.isprsjprs.2023.09.022

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  ChenY.ZhangL.ChenB.ZuoJ.HuY. (2024). MPG-Net: a semantic segmentation model for extracting aquaculture ponds in coastal areas from sentinel-2 MSI and planet SuperDove images. Remote Sens.16, 3760. 10.3390/rs16203760

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  DengC.HeR.WuZ.SunX.WangS. (2025). LU-Net: lightweight U-shaped network for water body extraction of remote sensing images. Water17, 2763. 10.3390/w17182763

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  DiakogiannisF. I.WaldnerF.CaccettaP.WuC. (2020). ResUNet-a: a deep learning framework for semantic segmentation of remotely sensed data. ISPRS J. Photogramm. Remote Sens.162, 94–114. 10.1016/j.isprsjprs.2020.01.013

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  DuanY.YangD.QuX.ZhangL.ChaoL.GanP.et al (2025). LCIRE-Net: lightweight cross-modal information interaction for road feature extraction from remote sensing images and GPS trajectory/LiDAR. IEEE Trans. Geosci. Remote Sens.63, 1–18. 10.1109/tgrs.2024.3516840

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  FayazM.NamJ.DangL. M.SongH.-K.MoonH. (2024). Land-cover classification using deep learning with high-resolution remote-sensing imagery. Appl. Sci.14, 1844. 10.3390/app14051844

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  FeyisaG. L.MeilbyH.FensholtR.ProudS. R. (2014). Automated water extraction index: a new technique for surface water mapping using landsat imagery. Remote Sens. Environ.140, 23–35. 10.1016/j.rse.2013.08.029

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  FisherA.FloodN.DanaherT. (2016). Comparing landsat water index methods for automated water classification in eastern Australia. Remote Sens. Environ.175, 167–182. 10.1016/j.rse.2015.12.055

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  GuoZ.WuL.HuangY.GuoZ.ZhaoJ.LiN. (2022). Water-body segmentation for SAR images: past, current, and future. Remote Sens.14, 1752. 10.3390/rs14071752

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  GuoB.ZhangJ.LiX. (2023). River extraction method of remote sensing image based on edge feature fusion. IEEE Access11, 73340–73351. 10.1109/access.2023.3296641

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  HanK.WangY.TianQ.GuoJ.XuC.XuC. (2020). “GhostNet: more features from cheap operations,” in Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, June 13–19, 2020, 1577–1586. 10.1109/CVPR42600.2020.00165

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  HongD.GaoL.YokoyaN.YaoJ.ChanussotJ.DuQ.et al (2021). More diverse means better: multimodal deep learning meets remote-sensing imagery classification. IEEE Trans. Geosci. Remote Sens.59, 4340–4354. 10.1109/TGRS.2020.3016820

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  HuangB.LiP.LuH.YinJ.LiZ.WangH. (2024). WaterDetectionNet: a new deep learning method for flood mapping with SAR image convolutional neural network. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.17, 14471–14485. 10.1109/JSTARS.2024.3440995

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  HuangL.JiangB.LvS.LiuY.FuY. (2024). Deep-learning-based semantic segmentation of remote sensing images: a survey. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.17, 8370–8396. 10.1109/JSTARS.2023.3335891

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  LeeC.LeeJ.KimT.LeeH.JavedA.ChungM.et al (2025). MultiVeg: a very high-resolution benchmark for deep learning-based multi-class vegetation segmentation. Remote Sens.18, 28. 10.3390/rs18010028

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  LiR.ZhengS.ZhangC.DuanC.WangL.AtkinsonP. M. (2021). ABCNet: attentive bilateral contextual network for efficient semantic segmentation of fine-resolution remotely sensed imagery. ISPRS J. Photogramm. Remote Sens.181, 84–98. 10.1016/j.isprsjprs.2021.09.005

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  LiK.JiH.LiZ.CuiZ.LiuC. (2025). AFNE-Net: semantic segmentation of remote sensing images via attention-based feature fusion and neighborhood feature enhancement. Remote Sens.17, 2443. 10.3390/rs17142443

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  LiuS.ChengJ.LiangL.BaiH.DangW. (2021). Light-weight semantic segmentation network for UAV remote sensing images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.14, 8287–8296. 10.1109/JSTARS.2021.3104382

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  LiuM.LiuJ.HuH. (2024). A novel deep learning network model for extracting lake water bodies from remote sensing images. Appl. Sci.14, 1344. 10.3390/app14041344

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  LvJ.ShenQ.LvM.LiY.ShiL.ZhangP. (2023). Deep learning-based semantic segmentation of remote sensing images: a review. Front. Ecol. Evol.11, 1201125. 10.3389/fevo.2023.1201125

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  MaL.LiuY.ZhangX.YeY.YinG.JohnsonB. A. (2019). Deep learning in remote sensing applications: a meta-analysis and review. ISPRS J. Photogramm. Remote Sens.152, 166–177. 10.1016/j.isprsjprs.2019.04.015

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  MiaoL.FengX.YangL.RenY.DengY.HangT. (2024). Spatiotemporal analysis of water body in the Chongming Island region over the past decade based on the ISUNet model. ISPRS Int. J. Geo-Inf.13, 134. 10.3390/ijgi13040134

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  MisbahM.KhanM. U.KaleemZ.MuqaibelA.AlamM. Z.LiuR.et al (2025). MSF-GhostNet: computationally efficient YOLO for detecting drones in low-light conditions. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.18, 3840–3851. 10.1109/JSTARS.2024.3524379

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  PoudelR. P. K.LiwickiS.CipollaR. (2019). “Fast-SCNN: fast semantic segmentation network,” in Proceedings of the British Machine Vision Conference, Cardiff, UK, September 9–12, 2019 (BMVC), 1–12.

  • Google Scholar
  • View reference in article

• 31

  RonnebergerO.FischerP.BroxT. (2015). “U-Net: convolutional networks for biomedical image segmentation,” in Proceedings of the 18th International Conference on Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015, Munich, Germany, 5–9 October, 2015, 234–241. 10.1007/978-3-319-24574-4_28

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  RostamiM.KolouriS.EatonE.KimK. (2019). Deep transfer learning for few-shot SAR image classification. Remote Sens.11, 1374. 10.3390/rs11111374

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  SarpG.OzcelikM. (2017). Water body extraction and change detection using time series: a case study of Lake Burdur, Turkey. J. Taibah Univ. Sci.11, 381–391. 10.1016/j.jtusci.2016.04.005

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  ShenX.WangH.WeiB.CaoJ. (2023). Real-time scene classification of unmanned aerial vehicles remote sensing image based on modified GhostNet. PLoS One18, e0286873. 10.1371/journal.pone.0286873

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  SongY.XueB.MengY.QinX.LiY.LiuQ. (2025). A fusion method incorporating dual-attention mechanism and transfer learning into UNet++ for remote sensing image coastline extraction. IEEE Access13, 11320–11331. 10.1109/access.2024.3467998

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  StoianA.PoulainV.IngladaJ.PoughonV.DerksenD. (2019). Land cover maps production with high resolution satellite image time series and convolutional neural networks: adaptations and limits for operational systems. Remote Sens.11, 1986. 10.3390/rs11171986

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  SunW.ChenC.LiuW.YangG.MengX.WangL.et al (2023). Coastline extraction using remote sensing: a review. GISci. Remote Sens.60, 2243671. 10.1080/15481603.2023.2243671

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  SunY.HuangL.ZhaoJ.LiX.QiuM. (2023). DSMFFNet: depthwise separable multiscale feature fusion network for bridge detection in very high resolution satellite images. Geocarto Int.38, 1–26. 10.1080/10106049.2022.2146761

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  ThirugnanasammandamoorthiP.GhoshD.DewanganR. K.HasanM. K.AriffinK. A. Z.AbbasH. S.et al (2025). FloodNet-Lite: a lightweight deep learning for flood mapping using remote sensing data with optimized UNet and edge deployment approach in 6G. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.18, 20294–20314. 10.1109/JSTARS.2025.3591406

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  TongX.-Y.XiaG.-S.LuQ.ShenH.LiS.YouS.et al (2020). Land-cover classification with high-resolution remote sensing images using transferable deep models. Remote Sens. Environ.237, 111322. 10.1016/j.rse.2019.111322

  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  TongQ.WuJ.ZhuZ.ZhangM.XingH. (2024). STIRUnet: swintransformer and inverted residual convolution embedding in unet for sea–land segmentation. J. Environ. Manage.357, 120773. 10.1016/j.jenvman.2024.120773

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  WangK. (2019). Evolution of Yellow River Delta Coastline based on remote sensing from 1976 to 2014, China. Chin. Geogr. Sci.29, 181–191. 10.1007/s11769-019-1023-5

  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  WangY.LiS.LinY.WangM. (2021). Lightweight deep neural network method for water body extraction from high-resolution remote sensing images with multisensors. Sensors21, 7397. 10.3390/s21217397

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  WangZ.LiC.WangX. (2021). “Convolutional neural network pruning with structural redundancy reduction,” in Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, June 20–25, 2021, 14908–14917. 10.1109/CVPR46437.2021.01467

  • CrossRef
  • Google Scholar
  • View reference in article

• 45

  WangL.LiR.ZhangC.FangS.DuanC.MengX.et al (2022). UNetFormer: a UNet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery. ISPRS J. Photogramm. Remote Sens.190, 196–214. 10.1016/j.isprsjprs.2022.06.008

  • CrossRef
  • Google Scholar
  • View reference in article

• 46

  WengL.XuY.XiaM.ZhangY.LiuJ.XuY. (2020). Water areas segmentation from remote sensing images using a separable residual SegNet network. ISPRS Int. J. Geo-Inf.9, 256. 10.3390/ijgi9040256

  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  WielandM.LiY.MartinisS. (2019). Multi-sensor cloud and cloud shadow segmentation with a convolutional neural network. Remote Sens. Environ.230, 111203. 10.1016/j.rse.2019.111203

  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  WooS.ParkJ.LeeJ.-Y.KweonI. S. (2018). “CBAM: convolutional block attention module,” in Proceedings of the 15th European Conference on Computer Vision—ECCV 2018, Munich, Germany, September 8–14, 2018, 3–19. 10.1007/978-3-030-01234-2_1

  • CrossRef
  • Google Scholar

• 49

  XiongX.WangX.ZhangJ.HuangB.DuR. (2023). TCUNet: a lightweight dual-branch parallel network for sea–land segmentation in remote sensing images. Remote Sens.15, 4413. 10.3390/rs15184413

  • CrossRef
  • Google Scholar

• 50

  YuC.GaoC.WangJ.YuG.ShenC.SangN. (2021). BiSeNet V2: bilateral network with guided aggregation for real-time semantic segmentation. Int. J. Comput. Vis.129, 3051–3068. 10.1007/s11263-021-01515-2

  • CrossRef
  • Google Scholar

• 51

  YuanY.WeiP.QiZ.DengX.ZhangJ.GanJ.et al (2025). Water body identification from satellite images using a hybrid evolutionary algorithm-optimized U-Net framework. Biomimetics10, 732. 10.3390/biomimetics10110732

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52

  ZhangZ.LiuQ.WangY. (2018). Road extraction by deep residual U-Net. IEEE Geosci. Remote Sens. Lett.15, 749–753. 10.1109/LGRS.2018.2802944

  • CrossRef
  • Google Scholar

• 53

  ZhangL.ZhangN.ShiR.WangG.XuY.ChenZ. (2023). SG-Det: Shuffle-GhostNet-based detector for real-time maritime object detection in UAV images. Remote Sens.15, 3365. 10.3390/rs15133365

  • CrossRef
  • Google Scholar

• 54

  ZhangT.LiW.FengX.RenY.QinC.JiW.et al (2024). “Super-resolution water body extraction based on MF-SegFormer,” in Proceedings of the IGARSS 2024—2024 IEEE International Geoscience and Remote Sensing Symposium, Athens, Greece, July 7–12, 2024, 9848–9852. 10.1109/IGARSS53475.2024.10640498

  • CrossRef
  • Google Scholar