<?xml version="1.0" encoding="utf-8"?>
    <rss version="2.0">
      <channel xmlns:content="http://purl.org/rss/1.0/modules/content/">
        <title>Frontiers in Remote Sensing | Image Analysis and Classification section | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/remote-sensing/sections/image-analysis-and-classification</link>
        <description>RSS Feed for Image Analysis and Classification section in the Frontiers in Remote Sensing journal | New and Recent Articles</description>
        <language>en-us</language>
        <generator>Frontiers Feed Generator,version:1</generator>
        <pubDate>2026-05-03T20:37:58.801+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2026.1812755</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2026.1812755</link>
        <title><![CDATA[S3ViT: self-supervised spectral vision transformer framework for hyperspectral unmixing]]></title>
        <pubdate>2026-04-22T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Dario Scilla</author><author>Victor Angulo</author><author>Kasper Johansen</author><author>Naif Alsalem</author><author>Wolfgang Heidrich</author><author>Matthew F. McCabe</author>
        <description><![CDATA[Hyperspectral unmixing aims to decompose each pixel in a hyperspectral image into a set of constituent endmembers and their corresponding abundances. Recent deep learning based approaches have demonstrated strong performance in capturing both spectral and spatial features. However, obtaining reliable per-pixel abundance ground truth in real hyperspectral scenes is generally infeasible, which motivates unsupervised and self-supervised unmixing strategies. In this work, we propose S3ViT, a self-supervised Spectral Vision Transformer designed for pixel-wise hyperspectral unmixing. The transformer captures spectral and spatial dependencies by applying self-attention over the full sequence of pixel tokens (1×1) augmented with learnable positional embeddings. It operates without ground-truth annotations by generating pseudo labels through an unsupervised process: first, Singular Value Decomposition (SVD) is used to estimate the number of endmembers based on a thresholded singular value spectrum; then, k-means clustering provides cluster-derived priors that are used to form a contextual token and initialize spectral prototypes, without being treated as true abundance supervision. To guide training, two initialization tokens, one from Vertex Component Analysis (VCA) and one from the k-means cluster-derived priors, are embedded alongside patch tokens into the transformer. The model learns to estimate abundance maps while enforcing the Abundance Non-negativity and Sum-to-One constraints through ReLU and Softmax layers. Endmember spectra are later estimated from pixels with high predicted abundances. We evaluated S3ViT on the Samson, Jasper Ridge and Washington DC Mall benchmark datasets and compared it to state-of-the-art geometrical and deep learning methods. Our model achieves superior or comparable performance in both RMSE and SAD metrics, with up to 31% improvement in SAD and 25% in RMSE. These results indicate that a compact pixel-token ViT, guided by weak spectral priors and optimized via reconstruction losses, can achieve competitive unmixing performance on standard benchmarks.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2026.1765013</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2026.1765013</link>
        <title><![CDATA[Benchmarking machine learning classifiers for urban mapping in arid environments: a google earth engine analysis of Riyadh’s expansion (1990–2025)]]></title>
        <pubdate>2026-03-24T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Amal Abdelsattar</author>
        <description><![CDATA[Monitoring urban expansion in arid regions is complicated by the spectral similarity between impervious surfaces and bare soil. Although machine learning classifiers on platforms like Google Earth Engine (GEE) offer effective solutions, their performance in these environments has not been systematically benchmarked. This study addresses this gap by comparing five supervised ML algorithms—Random Forest (RF), Support Vector Machine (SVM), Gradient Boosted Trees (GBT), Classification and Regression Tree (CART), and k-Nearest Neighbor (KNN)—for binary urban and non-urban mapping. We applied this analysis to Riyadh, Saudi Arabia, using a 35-year Landsat time series from 1990 to 2025. Annual, radiometrically consistent median composites were generated in GEE, with the 2025 composite based on imagery from January to September 2025. A custom ten-band feature stack, including indices such as the Bare Soil Index (BSI), was used for classification. The Random Forest model (RF-100) achieved the highest accuracy (Overall Accuracy = 0.977, Kappa = 0.954) and was selected for final land-use and land-cover mapping. Validation with 600 independent samples per epoch and comparison to the ESA WorldCover 2020 dataset confirmed the robustness of the results. The analysis found a 293% increase in Riyadh’s built-up area, from 416 km2 in 1990 to 1,219 km2 in 2025, with a notable slowdown in growth after 2010. Variable importance analysis showed that the Bare Soil Index (BSI) and the Normalized Difference Built-up Index (NDBI) were the most significant features for class separation, offering key methodological insights for arid urban remote sensing. This work provides a transferable methodological framework for classifier and feature selection in arid environments and a high-accuracy spatiotemporal dataset establishing a baseline for assessing sustainable urban development under Saudi Vision 2030.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2026.1751006</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2026.1751006</link>
        <title><![CDATA[A fully satellite-driven workflow for hydrodynamic modeling in data-scarce coastal systems: integrating ICESat-2, Sentinel-2, SWOT and reanalysis models]]></title>
        <pubdate>2026-03-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ali Reza Payandeh</author><author>Marc Simard</author><author>Daniel Jensen</author><author>Anthony Daniel Campbell</author><author>Heidi van Deventer</author><author>Alexandra Christensen</author>
        <description><![CDATA[Hydrodynamic models in coastal and estuarine systems are typically constrained by sparse bathymetry, boundary, and validation data, especially in regions where field campaigns are costly or impractical. Here we develop and test a fully satellite-driven framework for hydrodynamic modeling in South Africa’s Langebaan Lagoon without using any local in situ measurements. Bathymetry is derived by training multispectral Sentinel-2 reflectance against ICESat-2 ATL24 photon-derived depths using an XGBoost model optimized with Bayesian search. The final satellite derived bathymetry reproduces independent ATL24 points with RMSE = 0.45 m and R2 = 0.97. This bathymetry was used in a depth-averaged Delft3D Flexible Mesh model driven at the open boundary by TPXO tidal harmonics and by ERA5 winds. We validate modeled water surface elevation against 16 SWOT low-rate (250 m, unsmoothed) passes in 2023. SWOT–model comparisons yield an overall RMSE of 0.11 m and R2 = 0.61, with typical point differences <0.10 m (∼5% of the 2 m tidal range), and showed consistent spatial gradients in water level from the offshore boundary, through Saldanha Bay, and into the lagoon. At the offshore boundary, TPXO and SWOT sea surface heights agree closely (R2 = 0.86). A ∼26 min phase lag, determined using a lag-correlation analysis, reduces the TPXO–SWOT RMSE from 0.18 m to 0.11 m, indicating that phase differences explain some of the mismatch, with remaining differences likely linked to non-tidal signals. Our results demonstrate that combining passive optical, photon-counting LiDAR, radar interferometry, and global tidal/atmospheric models enables robust, transferrable hydrodynamic modeling in data-scarce coastal systems, offering a cost-effective pathway for monitoring.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1731775</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1731775</link>
        <title><![CDATA[Research on automatic mosaicking and synthesis processing technology for multi-source remote sensing images]]></title>
        <pubdate>2026-01-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jing Cai</author><author>Feng Ye</author><author>Jingyu Sun</author><author>Hangan Wei</author><author>Zichuang Li</author><author>Pengao Li</author>
        <description><![CDATA[Multi - source remote sensing image automatic mosaic and synthesis processing technology is the key to improving the utilization efficiency of remote sensing data. With the rapid develop-ment of diversified imaging platforms such as satellites, unmanned aerial vehicles and ground sensors, the heterogeneity of image data sources has become increasingly prominent, which makes the difficulty of mosaic and synthesis increase. This paper focuses on the auto-matic mosaic and synthesis processing technology of multi - source remote sensing images. Firstly, an adaptive block - weighted Wallis parallel color equalization algorithm fusing specific scene constraints is designed. It dynamically adjusts the block size of color equalization pro-cessing through the coefficient of variation, and optimizes the calculation of local color param-eters combined with bilinear interpolation, which avoids the color distortion of traditional glob-al algorithms and significantly improves the efficiency of radiometric correction. Moreover, an adaptive mosaic algorithm is introduced, and a space - constrained Markov Random Field - Graph Cut seamline generation model is used to generate seamless synthetic images, which supports large - area coverage. This technology can be extended to environmental monitoring, disaster assessment and urban planning. It can automatically process massive multi - source da-ta and achieve high - precision synthesis.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1678991</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1678991</link>
        <title><![CDATA[Geological mapping of copper deposits in the democratic republic of Congo through remote sensing data and machine learning]]></title>
        <pubdate>2026-01-14T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Matthieu Tshanga Matthieu</author><author>Lindani Ncube</author><author>Kgabo Humphrey Thamaga</author>
        <description><![CDATA[IntroductionThis study aimed to identify hydrothermal alteration zones and structural features associated with copper mineralisation in the Musonoi region, Lualaba Province, Democratic Republic of the Congo (DRC), within the Central African Copperbelt, using remote sensing and machine learning (ML). The study responds to the need for cost-effective and scalable exploration approaches in structurally complex tropical terrains.MethodsMultispectral satellite data from ASTER and Landsat 8 OLI, integrated with field observations and borehole information, supported the development of a predictive model. Principal Component Analysis (PCA), band ratios, and both manual and automated lineament extraction were used to enhance spectral and structural features. A lineament density map and hydrothermal alteration indices were produced and integrated with geological field data to verify the relationship between surface anomalies and subsurface mineralisation.ResultsRandom Forest classification indicated strong mineralisation in zones with high lineament density, fault intersections, and chlorite and kaolinite alteration. The main controlling variables were lineament density at 33.6%, fault proximity at 31.0%, and hydrothermal alteration index at 26.3%. Siliceous laminated rocks and basal dolomitic shale hosts mineralised units. Field validation confrmed that the model reflects known deposits, showing the strength of remote sensing and machine learning for exploration in complex tropical terrains.DiscussionThe study highlights the novelty of integrating Random Forest with multi source geospatial information in a structurally complex tropical terrain, and shows that this approach provides a cost effective and scalable tool for exploration in the Central African Copperbelt and similar geological provinces. Limitations related to spatial resolution and training data coverage remain and should be addressed in future work.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1718058</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1718058</link>
        <title><![CDATA[Efficient remote sensing image super-resolution with residual-enhanced wavelet and key-value adaptation]]></title>
        <pubdate>2026-01-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Rongchang Lu</author><author>Hongbo Miao</author><author>Xin Hai</author>
        <description><![CDATA[Remote sensing image super-resolution (SR) is vital for urban planning, precision agriculture, and environmental monitoring, yet existing methods have limitations: CNNs with restricted receptive fields cause edge blurring, Transformers with O(L2d) complexity fail in gigapixel-scale processing, and SSMs (e.g., Mamba) have directional biases missing diagonal features. To address these issues, this study proposes the REW-KVA architecture, integrating three innovations: Residual-Enhanced Wavelet Decomposition for separating low/high-frequency features and suppressing noise; Linear Attention with Key-Value Adaptation (complexity O(Ld)) for global context modeling; and Quad-Directional Scanning for omnidirectional feature capture. Validated on five datasets (DFC 2019, OPTIMAL-31, RSI-CB, WHU-RS19, UCMD), REW-KVA achieves state-of-the-art PSNR (29.17 dB on DFC 2019, 31.08 dB on RSI-CB) and SSIM (0.8958 on DFC 2019, 0.9442 on RSI-CB). It reduces memory reads by 35%, parameters by 42% (vs. SwinIR), and processes 1024×1024 images in 0.47 s (3.2× faster than SwinIR), serving as a deployable solution for resource-constrained platforms.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1697897</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1697897</link>
        <title><![CDATA[Enhancing land cover classification in the heterogeneous landscape by integrating auxiliary data with Sentinel-2 imagery using the random forest algorithm]]></title>
        <pubdate>2026-01-09T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Irvin D. Shandu</author><author>Sifiso Xulu</author><author>Michael Gebreslasie</author>
        <description><![CDATA[Effective and accurate land use and land cover classification (LULC-C) is an indispensable exercise for various environmental management objectives, including past and future land-use dynamics, flood and runoff modelling. However, LULC-C is subject to several limitations, such as labour-intensive derivation of a labelled dataset. So, we aim to enhance LULC-C using auxiliary features: elevation, slope, aspect, distance from-(road, railway stations, rivers, water, and town), global human settlement built-up layer and remote sensing indices in the heterogeneous landscape of eThekwini (EM) and Nelson Mandela Bay Metropolitan (NMBM) using Sentinel-2 and random forests (RF). We compared two classification scenarios: (1) feature set including bands and indices, and (2) feature set including bands, indices, and auxiliary features. We trained and tested RF using block cross-validation and random hold-out (70/30) split and validated the classified image using independent validation and 30% subset, through overall accuracy (OA) and F1-score. The study quantified the uncertainty using a 95% confidence interval with bootstrapping samples of 1000 iterations, and quantify the significance of scenario 2 using McNemar and p-value. Pixel-wise quantity and allocation disagreement were derived to compare classification scenarios against the two 2020 reference maps for South African National Land Cover and Environmental System Research Institute. A class-by-class pixel comparison between classification scenarios underscores the potential of auxiliary features. While classification scenarios achieved comparable accuracy, scenario 2 superseded scenario 1 in all classification scheme. Using an independent validation, the study found confidence interval (CI) for OA of 83.63% CI: 77.78–88.89 improved to 89.47% CI: 84.79–94.15, respectively, for scenario 1 and scenario 2 over EM. Confirmed by NMBM, where OA of 82.29% CI: 76.57–87.43 stabilised to 88.57% CI: 84.00–93.14 for scenario 1 and scenario 2. The performance improvement was statistically significant, attaining p-values of 0.03 and 0.02, respectively, for EM and NMBM using an independent validation set. However, while using 30% validation subset, results show in-significant improvement in NMBM attaining p-value = 0.07, where p-value >0.05. Overall results proved that an integration of auxiliary features enhance LULC-C. The empirical result of this study is a step forward in effective LULC-C in a complex landscape.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1693286</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1693286</link>
        <title><![CDATA[Detection potential of floating matter in high-resolution X-band SAR data: initial results with visual interpretations]]></title>
        <pubdate>2025-12-16T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Madjid Hadjal</author><author>Brian B. Barnes</author><author>Chuanmin Hu</author><author>Lin Qi</author><author>Dimitris Papageorgiou</author><author>Konstantinos Topouzelis</author>
        <description><![CDATA[Remote detection of floating matter, such as macroalgae, plastics, or other debris, primarily relies on the use of passive optical imagery that requires daytime collection and an absence of clouds, sun glint, and thick aerosols. Synthetic aperture radar (SAR) sensors are not affected by these issues, but their capacity in such detection has not been robustly characterized. As such, this study qualitatively evaluates the capacity of Capella Space X-band (9.6 GHz) SAR, which provides data at a spatial resolution of 0.35–1 m, 100 to 800 times higher than what is currently available from Sentinel-1 C-band (5.4 GHz) SAR. A controlled experiment with floating plastic targets of 1 m2, 2 m2, and 3 m2 showed that only the 3 m2 target and 1 m2 mooring buoys were clearly detected and only in a single collection mode. Some macroalgae and floating plants, such as Ulva prolifera and hyacinth, were consistently detected by Capella SAR. However, Sargassum horneri and Sargassum natans/fluitans were only partially detected by Capella SAR, with larger aggregations providing more positive detections. Surface scums of phytoplankton such as Trichodesmium or Noctiluca were not detected. The main detection limitations arise from the weak contrast between the floating matter and the surrounding water, as well as the low signal-to-noise ratios (SNRs) of the three different collection modes of Capella SAR, which range from 2 to 6 (±0.03–0.35). On the other hand, Capella SAR successfully detected floating material in Lake Skadar/Shkodra (Albania and Montenegro) and foam and potential brine shrimp cysts in the Great Salt Lake, while these targets were not detected using Sentinel-1. Despite a limited dataset of only 33 SAR images paired with concurrent and co-located optical images, these preliminary results show the value of high-resolution X-band SAR in detecting relatively large mats of plastics and certain types of macroalgae. The findings can also help task the satellites to collect targeted images for event response.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1678882</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1678882</link>
        <title><![CDATA[A privacy-preserving, on-board satellite image classification technique incorporating homomorphic encryption and transfer learning]]></title>
        <pubdate>2025-12-10T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Abhijit Roy</author><author>Mahendra Kumar Gourisaria</author><author>Rajdeep Chatterjee</author><author>Amitkumar V. Jha</author><author>Bhargav Appasani</author><author>Nicu Bizon</author><author>Alin Gheorghita Mazare</author>
        <description><![CDATA[Satellite image classification is an important and challenging task in the modern technological age. Satellites can capture images of danger-prone areas with very little effort. However, the size and number of satellite images are very high when they are rapidly captured from space, and they require a huge amount of memory to store the data. In addition, keeping the satellite images private is another important task for security purposes. On-board, instant, accurate classification of a smaller number of satellite images is a challenging task, which is important to determine the specific condition of an area for instant monitoring. In the proposed hybrid approach, the captured images are kept secure, while the required training of the classification is done separately. Finally, the trained module is encrypted for use by the satellite to perform the on-board classification task. The Brakerski–Fan–Vercauteren (BFV)-based homomorphic encryption of EuroSAT satellite images is applied to store images in a cloud storage, where the privacy of the images can be maintained. Then, the decrypted images are used for training four transfer learning models (YOLOv8, YOLOv12, ResNet34, ResNet101, and a vision transformer classification. The best-trained module is encoded and encrypted again by using homomorphic encryption to limit the module to authorized devices. The encrypted module is decrypted and decoded to recover the trained module, which is used for instant classification of test images. Finally, the performance of the transfer learning models is evaluated from the test results. The vision transformer classifier achieved the highest accuracy of 99.65%.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1680353</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1680353</link>
        <title><![CDATA[3D colored point cloud classification of a deep-sea cold-water coral and sponge habitat using geometric features and machine learning algorithms]]></title>
        <pubdate>2025-12-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Salem Morsy</author><author>Ana Belén Yánez-Suárez</author><author>Katleen Robert</author>
        <description><![CDATA[Classification of benthic habitats in the deep sea is instrumental in managing and monitoring marine ecosystems as it provides distinct units for which changes can be quantified over time. These applications require automatic classification approaches with reasonable accuracy to ensure efficiency and robustness. The use of 3D point clouds is currently emerging in deep-sea benthic classification as it allows for high-resolution representation of the 3D structure (i.e., geometry), texture, and composition of complex benthic habitats such as those created by structure-forming cold-water corals. Point clouds were derived from remotely operated vehicle video surveys of three vertical walls (depth range 1400–1900 m) along the Charlie-Gibbs Fracture Zone, North Atlantic. In addition to RGB values, this research incorporated nine geometric features derived from structure-from-motion 3D point clouds to classify coral and sponge colonies. Three unsupervised (k-means (KM), fuzzy c-means (FCM), and Gaussian mixture model (GMM)) and three supervised (decision tree (DT), random forest (RF), and linear discriminant analysis (LDA)) machine learning (ML) algorithms were compared and assessed for accuracy and reliability. The ML classifiers were used to build full-coverage seafloor predictions for three classes, namely, seabed, sponges, and corals. The KM, GMM, and FCM achieved an average overall accuracy of 74.87%, 71.94%, and 70.77%, respectively, while the RF, LDA, and DT achieved 84.50%, 84.01%, and 79.90%, respectively. Overall, the supervised ML classifiers outperformed the unsupervised ML classifiers. In particular, the RF classifier demonstrated the highest overall classification accuracy and F1-score for individual classes, with an average of 89.09%, 67.12%, and 41.60% for the seabed, sponges, and corals, respectively. In addition, the spatial coherence of the point clouds was considered and improved the results’ overall accuracy and F1-score by up to 9% and 12%, respectively. Results showed that incorporating geometric features, traditionally employed in terrestrial and shallow-water LiDAR surveys, in combination with RGB values is suitable for high-resolution deep-sea benthic 3D point clouds classification.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1696570</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1696570</link>
        <title><![CDATA[Coffee extraction from remote sensing imagery based on multiple features: a case study of Pu’er City, China]]></title>
        <pubdate>2025-11-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Qianrui Huang</author><author>Xianfeng Cheng</author><author>Yu Chen</author><author>Xinyi Ding</author><author>Huicong Jia</author>
        <description><![CDATA[IntroductionCoffee, a vital beverage and cultural symbol, significantly influences global economic and cultural development. Due to the characteristics of agricultural production activities, such as areas, significant differences, and relatively low economic benefits per unit area, Traditional ground surveys often fail to accurately capture coffee crop distribution due to the large-scale, regionally varied, and economically modest nature of agricultural production. Remote sensing offers a promising alternative but faces challenges in distinguishing coffee from vegetation with similar spectral characteristics, especially in areas with complex land cover and dense canopies.MethodsThis study focuses on Pu’er City in Yunnan Province, China, renowned as the ‘golden belt’ of global coffee cultivation. Using Sentinel-2 remote sensing imagery, we analyzed key phenological features through time-series curves of the Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), and Difference Vegetation Index (DVI). To ensure a balanced and representative dataset, interpretation keys were established from 1,617 field-measured sampling points, yielding a total of 4,000 coffee and non-coffee samples. Employing the Random Forest (RF) algorithm, we constructed a refined coffee crop extraction model incorporating spectral, texture, terrain, and regional pattern features.ResultsThe findings indicate: (1) Incorporating administrative division features and using a larger texture window size (5 × 5) enhances model accuracy, achieving an overall accuracy (OA) of 93.92% and a Kappa coefficient of 0.8783. (2) The four-period segmentation approach significantly improved accuracy, with the highest OA reaching 94.80%, identifying October to December (coffee fruiting season) as the most critical period for classification. (3) Administrative Division Features (ID), Topographical features (SLOPE) and vegetation indices (NDVI and DVI) were the most crucial for coffee classification, while texture features, except for Sum Average (SAVG), generally had lower importance.DiscussionThis study validates the effectiveness of remote sensing in monitoring and mapping coffee cultivation. The proposed feature input strategy shows strong potential for application in other regions with similar agro-ecological conditions, supporting precision agricultural management and promoting sustainable coffee farming practices.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1661528</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1661528</link>
        <title><![CDATA[Operationalizing remote sensing methods for smallholder dry season irrigation detection in sub-Saharan Africa]]></title>
        <pubdate>2025-10-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Hasan Siddiqui</author><author>Christopher Small</author><author>Vijay Modi</author>
        <description><![CDATA[In many parts of the tropics a prolonged dry season presents an economic opportunity for farmers to grow a second crop beyond an otherwise single crop that a shorter rainy season permits. These additional second crops can ensure food security, improve nutrition and increase incomes. The first contribution of this paper is to granularly identify regions of Sub-Saharan Africa where a prolonged dry season exists. Energy planners are also keen to assess where dry-season agriculture is being currently practiced and the extent of the area cropped in the dry season. Assuming this is carried out using irrigation, this allows planners to assess the scale of water and energy needs if these practices are to be scaled. The phenological characterization of the landscape using vegetation patterns helps to identify regions where dry season irrigation is feasible. This study operationalizes an irrigation detection methodology originally applied to the Ethiopian highlands built using visually collected labels from high resolution imagery and limited ground truth data. The second contribution of the paper lies in the application of the methodology over a range of African geographies, with the exclusive use of visually collected labels. The methodology relies on the distinct phenology of irrigated crops in the dry season that differentiates them from rain-fed agriculture and evergreen vegetation. The method is applied across different countries in sub-Saharan Africa to detect smallholder plots that are as small as a tenth of a hectare. The method is found to be viable in semi-arid areas with a prolonged dry season such as Northern Nigeria and Burkina Faso. We demonstrate how humid regions such as those in Uganda with longer duration rainfall are not well suited for the methodology. This is because the short dry season does not allow sufficient time for non-irrigated vegetation to senesce making it difficult to distinguish dry-season irrigation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1685140</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1685140</link>
        <title><![CDATA[AU-super: superpixel scale optimization and training data augmentation strategy for hyperspectral image classification]]></title>
        <pubdate>2025-10-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Yiran Wang</author><author>Lingling Li</author><author>Mingming Xu</author><author>Shanwei Liu</author><author>Muhammad Yasir</author><author>Manuel Á. Aguilar</author><author>Fernando J. Aguilar</author>
        <description><![CDATA[The spectral information of each pixel in hyperspectral images contains valuable information about object properties, although accurate labeling is required in supervised classification to guide the model in distinguishing different land cover types. However, labeling data for hyperspectral images is difficult to obtain, especially in complex or remote areas. This results in a shortage of labeled samples, which prevents the model from fully learning the features of different classes. To overcome this challenge, this work proposes a hyperspectral image classification method, called AU-Super, that combines adaptive superpixel scale selection, superpixel label expansion, and data augmentation. First, an adaptive method is developed to determine an appropriate superpixel segmentation scale based on feature values, thereby ensuring that superpixel segmentation effectively captures the spatiospectral information of the image. Second, feature extraction is performed at the previously estimated superpixel scale. Third, pixel labels are converted to superpixel labels to reduce the effects of labeling noise during the training process. Furthermore, superpixel-level label-based data augmentation techniques are introduced to mitigate the problem of under-labeled patterns. The comparative results against various state-of-the-art algorithms demonstrate that AU-Super-RF consistently achieves superior performance across multiple accuracy metrics. Under few-shot training scenarios (with only 1–10 samples per class) on the Indian Pines, Salinas, and Pavia University datasets, it improves overall accuracy by 3%–7%, average accuracy by 2%–6%, and the Kappa coefficient by 3%–8%, highlighting the robustness and practical utility of the method.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1637820</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1637820</link>
        <title><![CDATA[Multi‐Scale graph wavelet convolutional network for hyperspectral image classification]]></title>
        <pubdate>2025-10-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Hao Zhang</author><author>Junhua Ku</author><author>Jie Zhao</author>
        <description><![CDATA[Hyperspectral images (HSIs) have very high dimensionality and typically lack sufficient labeled samples, which significantly challenges their processing and analysis. These challenges contribute to the dimensionality curse, making it difficult to describe complex spatial relationships, especially those with non-Euclidean characteristics. This paper presents a multi-scale graph wavelet convolutional network (MS-GWCN) that utilizes a graph wavelet transform within a multi-scale learning framework to accurately capture spatial-spectral features. The MS-GWCN constructs graphs according to 8-neighborhood connectivity schemes, implements spectral graph wavelet transforms for multi-scale decomposition, and aggregates features through multi-scale graph convolutional layers. Our method, the MS-GWCN, demonstrates superior performance compared to existing methodologies. It achieves higher overall accuracy, average accuracy, per-class accuracy, and the Kappa coefficient, as evaluated on three datasets, including the Indian Pines, Salinas, and Pavia University datasets, thereby demonstrating enhanced robustness and generalization capability.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1682132</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1682132</link>
        <title><![CDATA[Sentinel-2 forest typology mapping in Central Africa: assessing deep learning and image preprocessing effects]]></title>
        <pubdate>2025-09-26T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Gaëlle Viennois</author><author>Hadrien Tulet</author><author>Paul Tresson</author><author>Pierre Ploton</author><author>Pierre Couteron</author><author>Nicolas Barbier</author>
        <description><![CDATA[IntroductionCentral African forests are key reservoirs of carbon and biodiversity. Developing a detailed, spatially explicit typology of forest types is essential for monitoring and conservation. However, this task remains challenging due to limitations inherent to optical satellite imagery, especially disturbances caused by two major sources of noise: (i) atmospheric effects and (ii) Bidirectional Reflectance Distribution Function (BRDF) distortions, which introduce spectral inconsistencies across image collections. Even after standard corrections, residual errors often persist, masking the subtle ecological signals required for accurate classification. In this study, we evaluate whether recent deep learning models can implicitly learn to account for such distortions, potentially reducing the need for traditional preprocessing steps.MethodsWe produced a 10-m resolution vegetation typology map of the highly heterogeneous TRIDOM landscape (∼180,000 km2) spanning Cameroon, Gabon, and the Republic of Congo, using Sentinel-2 imagery. We compared the performance of Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and self-supervised ViTs trained with DINOv2.ResultsOur results show that CNNs achieved the highest classification accuracy (OA = 0.91, Kappa = 0.84), outperforming both ViTs and DINOv2-based models (OA ≈ 0.70) on preprocessing images. When uncorrected imagery was used, CNN accuracy dropped to 0.76 (Kappa = 0.59), while ViTs exhibited also a decline (Kappa falling from 0.54 to 0.24).DiscussionThese findings highlight the partial ability of deep learning models to compensate for image noise, but emphasize that traditional preprocessing remains necessary for reliable classification. Our results also demonstrate that CNNs consistently outperform self-supervised Vision Transformers in large-scale forest mapping, providing accurate classification of forest typologies. This work offers new insights into the robustness and current limitations of deep learning architectures when applied to complex tropical landscapes.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1668978</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1668978</link>
        <title><![CDATA[HPLNet: a hierarchical perception lightweight network for road extraction]]></title>
        <pubdate>2025-09-26T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Shilin Cui</author><author>Qi Feng</author><author>Luyan Ji</author><author>Xiaowen Liu</author><author>Baofeng Guo</author>
        <description><![CDATA[With the progression of remote sensing technologies, extracting road networks from satellite imagery has emerged as a pivotal research domain in both Geographic Information Systems and Intelligent Transportation Systems. Recognizing the difficulty in balancing lightweight network design with extraction accuracy, the challenge of synergistically preserving global road connectivity and local details, and the hardship in effectively integrating low-level features with high-level representations to achieve full coupling between road details and semantic understanding in road extraction from remote sensing images, this study introduces a Hierarchical Perception Lightweight Network for road extraction (HPLNet). This innovative network integrates shallow perception part and deep perception part, aiming to optimize the trade-off between inference efficiency and extraction accuracy. In shallow perception, directional stripe convolutions capture road details, while deep perception integrates a spatial-channel semantic awareness network to bridge local and global information, boosting road semantic feature extraction. Moreover, to extend the model’s reception at both pixel and semantic levels, each network component strategically introduces parameter-free channel shift operations. HPLNet attains state-of-the-art efficiency in balancing parameter footprint and inference latency: its parameter count is merely 22% of that of U-Net, while its inference speed is 18% faster than FCN. Concurrently, it delivers competitive segmentation metrics on the Massachusetts dataset, achieving an IoU of 64.32% and an F1 score of 79.96%. Experimental results demonstrate that the proposed network achieves superior performance in both segmentation accuracy and model complexity, thereby offering an efficient solution for real-time deployment on edge devices.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1662024</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1662024</link>
        <title><![CDATA[CARP: cloud-adaptive robust prompting of vision-language models for ship classification under cloud occlusion]]></title>
        <pubdate>2025-09-17T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Haoke Zhan</author><author>Yiping Song</author><author>Xun Huang</author><author>Xiao Tan</author><author>Ting Zhang</author>
        <description><![CDATA[Fine-grained few-shot ship classification under cloud occlusion is vital for maritime safety but remains challenging due to corrupted features and limited data utility. While the advent of large pre-trained vision-language models (VLMs) provides promising solutions, the lack of specialized benchmarks hinders their effective application. To address this, we introduce SeaCloud-Ship, the first benchmark dedicated to this task. It comprises 7,654 high-resolution, high-quality annotated images across 30 classes, featuring quantified cloud coverage (12.5%–75%) for standardized evaluation. We innovatively propose CARP, a cloud-aware prompting framework built upon CoOp, to combat feature corruption, semantic misalignment, and utility decay. Our core contributions include: (1) GCE Loss dynamically adjusting classification weights to suppress cloud interference based on feature degradation severity; (2) Adaptive Optimization Prompt Design (AOPD) utilizing distortion-aware vectors for effective multi-modal feature alignment and semantic deviation repair; (3) Dynamic Weight Adjustment Mechanism (DWAM) real-time balancing of multi-source feature fusion by evaluating inter-modal information gain. Extensive experiments on SeaCloud-Ship demonstrate CARP’s superior robustness and state-of-the-art performance, establishing a strong baseline for cloud-occluded ship classification.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1578841</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1578841</link>
        <title><![CDATA[A deep-learning framework to detect green tide from MODIS images]]></title>
        <pubdate>2025-09-10T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Weidong Zhu</author><author>Yuelin Xu</author><author>Lei Zhang</author><author>Zitao Liu</author><author>Shuai Liu</author><author>Yifei Li</author>
        <description><![CDATA[IntroductionMonitoring Ulva prolifera blooms over the long term is crucial for maintaining marine ecological balance. MODIS images, with their wide spatial coverage, high temporal resolution, and rich historical data, are commonly used for this purpose. However, their relatively low spatial resolution may lead to inaccuracies in precisely defining the bloom extents, thereby impeding the formulation of effective management strategies.MethodsTo address this issue, our study developed the WaveNet model. This model integrates VGG16 with the Bidirectional Feature Pyramid Network (BiFPN) and is further enhanced with a Convolutional Block Attention Module (CBAM). We applied this framework to MODIS imagery for the detection and monitoring of U. prolifera.ResultsWaveNet demonstrated superior performance in long-term sea surface U. prolifera monitoring compared to traditional methods, achieving an accuracy of 97.14% and an F1 score of 93.26%. This represents a significant improvement over existing techniques.DiscussionThese results highlight WaveNet’s improved capacity for accurate spatial recognition and classification, overcoming the limitations of previous methods. Applying this approach, we analyzed the spatiotemporal distribution of U. prolifera blooms in the Yellow Sea of China from 2018 to 2024. Our framework offers valuable insights for early prevention and targeted management of green tides, contributing to the development of more effective mitigation strategies.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1570827</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1570827</link>
        <title><![CDATA[Combination of neural network models for estimating Chlorophyll-a over turbid and clear waters (CONNECT)]]></title>
        <pubdate>2025-09-01T00:00:00Z</pubdate>
        <category>Methods</category>
        <author>Manh Duy Tran</author><author>Vincent Vantrepotte</author><author>Roy El Hourany</author><author>Daniel Schaffer Ferreira Jorge</author><author>Milton Kampel</author><author>João Felipe Cardoso dos Santos</author><author>Eduardo Negri Oliveira</author><author>Rodolfo Paranhos</author><author>Cédric Jamet</author>
        <description><![CDATA[Estimation of Chlorophyll-a concentration (Chl-a) across diverse aquatic systems using Moderate Resolution Imaging Spectroradiometer-Aqua (MODIS-A) data has posed challenges, particularly the inability of existing algorithms to maintain consistent accuracy across varying optical water conditions, from oligotrophic clear waters to highly turbid productive systems. Traditional Blue/Green ratio approaches often show limitations over optically complex waters where colored dissolved organic matter and suspended sediments interfere with phytoplankton signal detection. In contrast, Red/NIR (Near-Infrared) models perform relatively well in productive coastal domains but are less effective in open ocean waters where phytoplankton absorption is too weak to produce detectable signals in these longer wavelengths. To address these challenges, we developed a Combination Of Neural Network models for Estimating Chlorophyll-a over Turbid and clear waters (CONNECT model) based on the principle that different Optical Water Types (OWTs) require specialized bio-optical algorithms. The methodology involves the development of two Multi-Layer Perceptron (MLP) models (NN-Clear & NN-Turbid) that are trained and evaluated on a comprehensive in-situ dataset with simultaneous measurements of Remote Sensing Reflectance (Rrs) and Chl-a gathered in various environments from clear to ultra-turbid waters (N = 5,358) with Chl-a ranging between 0.017 and 838.24 µg.L-1. These specialized models are then combined through a weighted blending approach to produce unified Chl-a estimates that adapts to the optical conditions of various water types. In particular, the algorithm merging process involves the use of probability values corresponding to 2 groups of Optical Water Types as the blending coefficients. Accuracy evaluations performed on both in-situ and matchup datasets indicate a remarkable advancement of the CONNECT model compared to the traditional Blue/Green approaches over different trophic conditions with an improvement of 49.65% on the matchup validation considering the Symmetric Signed Percentage Bias (SSPB) metric.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frsen.2025.1599099</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frsen.2025.1599099</link>
        <title><![CDATA[Efficient vision transformers with edge enhancement for robust small target detection in drone-based remote sensing]]></title>
        <pubdate>2025-07-22T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Xuguang Zhu</author><author>Zhizhao Zhang</author>
        <description><![CDATA[Small object detection in UAV remote sensing imagery faces significant challenges due to scale variations, background clutter, and real-time processing requirements. This study proposes a lightweight transformer-based detector, MLD-DETR, which enhances detection performance in complex scenarios through multi-scale edge enhancement and hierarchical attention mechanisms. First, a Multi-Scale Edge Enhancement Fusion (MSEEF) module is designed, integrating adaptive pooling and edge-aware convolution to preserve target boundary details while enabling cross-scale feature interaction. Second, a Layered Attention Fusion (LAF) mechanism is developed, leveraging spatial depth-wise convolution and omnidirectional kernel feature fusion to improve hierarchical localization capability for densely occluded targets. Furthermore, a Dynamic Positional Encoding (DPE) module replaces traditional fixed positional embeddings, enhancing spatial perception accuracy under complex geometric perspectives through learnable spatial adapters. Combined with an Inner Generalized Intersection-over-Union (Inner-GIoU) loss function to optimize bounding box geometric consistency, MLD-DETR achieves 36.7% AP50% and 14.5% APs on the VisDrone2019 dataset, outperforming the baseline RT-DETR by 3.2% and 1.8% in accuracy while achieving 20% parameter reduction and maintaining computational efficiency suitable for UAV platforms equipped with modern edge computing hardware. Experimental results demonstrate the algorithm’s superior performance in UAV remote sensing applications such as crop disease monitoring and traffic congestion detection, offering an efficient solution for real-time edge-device deployment.]]></description>
      </item>
      </channel>
    </rss>