Remote Sensing Applications-Society and Environment最新文献

{"title":"Geospatial impact evaluation of an agricultural intensification program","authors":"Fiona Kastel,&nbsp;Sanchi Lokhande,&nbsp;Devika Lakhote,&nbsp;Anca Dumitrescu,&nbsp;Douglas Glandon","doi":"10.1016/j.rsase.2026.102014","DOIUrl":"10.1016/j.rsase.2026.102014","url":null,"abstract":"<div><div>Climate change is exacerbating food insecurity worldwide, particularly in drought-prone regions, such as the Sahel region of West Africa. To improve food security and resilience, the government of Niger, with funding from the West African Development Bank, implemented a multi-faceted agricultural production intensification program. Evaluating the effectiveness of such programs over time is challenging in remote and conflict-affected regions. We use remote sensing data with a novel econometric design to measure the impact of constructing surface irrigation infrastructure and land rehabilitation projects within this program on agricultural production and water retention, using vegetation indices (Normalized Difference Vegetation Index and Soil Adjusted Vegetation Index) and soil moisture bands, respectively. We employ 500 m² pixel-level remote sensing data from Landsat-7, TerraClimate, and CHIRPS. We measure the direct impact of irrigation and land rehabilitation on those who benefited and the overall potential impact of providing surface irrigation on all intended recipients, whether or not they used it, to assess agricultural production outcomes. Results show that there are significant increases in vegetation and soil moisture due to the surface irrigation, particularly in the dry season. We also find heterogeneous impacts by project region, with a larger increase in vegetation and soil moisture in the Ibohamane village than the Dougeruoua village. At the same time, the effect of the combined land rehabilitation interventions on agricultural production and soil moisture are mixed, and ultimately, negligible. We demonstrate the unique value that using a synthetic difference-in-difference design alongside open-source remotely sensed data can provide to measure the long-term impact of agricultural interventions, particularly irrigation projects, which tend to fail over time in certain contexts.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 102014"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147710212","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"BGR-net: A boundary-guided refinement network for multi-class cloud detection","authors":"Shaowei Bai ,&nbsp;Yan Mo ,&nbsp;Wanting Zhou","doi":"10.1016/j.rsase.2026.102024","DOIUrl":"10.1016/j.rsase.2026.102024","url":null,"abstract":"<div><div>To address the limitations of existing cloud detection methods in handling blurred boundaries, omission of fragmented small-scale clouds, and interference from complex terrestrial backgrounds, this paper proposes a Boundary-Guided Refinement Network (BGR-Net). The network adopts a hierarchical architecture that deeply integrates the local inductive bias of Convolutional Neural Networks (CNNs) with the global long-range dependency modeling capabilities of Transformers. During the encoding stage, a multi-scale downsampling attention mechanism and a convolutional positional enhancement module are employed to dynamically capture rich spatial details and content-aware information across scales. In the decoding stage, a cross-layer modulation mechanism effectively integrates multi-level features from the encoder and decoder, significantly enhancing the discriminative power of cloud-related features while suppressing background noise. Furthermore, by leveraging cloud boundary and positional feature guidance, the network achieves efficient reconstruction of complex cloud edges. Experimental results based on the Himawari-9 satellite remote sensing dataset demonstrate that BGR-Net strikes a superior balance between segmentation accuracy and computational efficiency. Its mIoU (68.37%) and F1-score (81.21%) significantly outperform current state-of-the-art models. With its low computational complexity, BGR-Net fulfills the practical requirements for real-time monitoring in remote sensing applications. Code is available at: <span><span>https://github.com/NotRoots/BGR-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 102024"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147710291","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"An analysis of the regenerative capacity of vegetation using NDVI and NBR after large-scale wildfires (1990-2024) in the Mediterranean province of Castellón, Spain","authors":"Enrique Montón Chiva,&nbsp;José Quereda Sala","doi":"10.1016/j.rsase.2026.102020","DOIUrl":"10.1016/j.rsase.2026.102020","url":null,"abstract":"<div><h3>Purpose</h3><div>This study analyses one of the ways in which Mediterranean vegetation adapts to climate and wildfires: its regenerative capacity following wildfires that occurred in the Spanish province of Castellón. It also evaluates the suitability of satellite indices for assessing this regenerative capacity using field photographs and Google Earth imagery.</div></div><div><h3>Methods</h3><div>The study examines wildfire databases in Spain and Castellón and changes in NDVI (Normalized Difference Vegetation Index) between 1990 and 2024 in the affected areas using Landsat images processed with the Climate Engine application, as well as NBR (Normalized Burn Ratio) values calculated from Landsat imagery in GEE (Google Earth Engine). Both indices are calculated using distinct spectral bands from the Landsat satellite series. This regeneration capacity is quantified by comparing the values recorded from pre-fire and immediately post-fire conditions, in addition to analysing the indices five years after the fire event.</div></div><div><h3>Results</h3><div>The analysis reveals a downward trend in fires in Spain and Castellón, especially in the number of fire events. The NDVI values also indicate the strong regenerative capacity of well-adapted Mediterranean vegetation, with values rising to above the threshold of 75% of the pre-fire NDVI in all cases within five years. Conversely, the NBR percentages are less impressive.</div></div><div><h3>Conclusions</h3><div>The variations observed in both indices only partially reflect the actual situation since, as field photographs and the Google Earth application show, although the vegetation cover has recovered, its structure has not, as this process takes a longer period of time.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 102020"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147710293","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"GNSS-IR and spectral remote sensing data fusion for soil moisture estimation","authors":"R. Awad ,&nbsp;F. Kizel ,&nbsp;G. Even-Tzur","doi":"10.1016/j.rsase.2026.101983","DOIUrl":"10.1016/j.rsase.2026.101983","url":null,"abstract":"<div><div>Recent advancements in Global Navigation Satellite System-Interferometric Reflectometry (GNSS-IR) have enabled the extraction of environmental data by analyzing differences between direct and multipath signals. A key application is soil moisture estimation using Signal-to-Noise Ratio (SNR) measurements of reflected signals. However, these methods only provide relative moisture estimates, requiring periodic in-situ measurements to establish the minimum moisture level for each observation period. On the other hand, optical remote sensing estimates soil moisture through pixel reflectance, clearing the need for in-situ measurements but is limited by sensitivity to weather and illumination conditions, cloud cover, ground vegetation cover, and a 3–5-day satellite orbit, hindering continuous estimation. This study further develops the potential of GNSS-IR for soil moisture estimation by introducing a novel, optimized approach to enhance accuracy. We also develop a data fusion model that combines GNSS-IR's continuous, weather-independent measurements with discrete estimates from spectral remote sensing Sentinel-2 imagery. This model enables continuous soil moisture estimation without in-situ measurements. We use datasets from Valencia, Spain, and Kabri, Israel, to evaluate the methodology. Our models achieve an accuracy relative to in-situ data of approximately 0.02 [m³/m³] in soil moisture estimation, outperforming traditional methods, which have an accuracy of around 0.05 [m³/m³].</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 101983"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147613418","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Early discrimination of conservation tillage practices in wheat using a novel spectral index from Sentinel-2 time series","authors":"Rajkumar Dhakar ,&nbsp;Vinay Kumar Sehgal ,&nbsp;Parmod Kumar ,&nbsp;Niveta Jain","doi":"10.1016/j.rsase.2026.101967","DOIUrl":"10.1016/j.rsase.2026.101967","url":null,"abstract":"<div><div>Monitoring, Reporting, and Verification (MRV) of regenerative agricultural practices such as conservation tillage is crucial for credible carbon markets and sustainable food systems in India's rice-wheat belt. Addressing this need, we developed a novel, satellite-based approach to discriminate between conventional tillage (CT) and conservation tillage [Happy Seeder (HS) and Super Seeder (SS)] practices of wheat sowing early in the season (during sowing window) using Sentinel-2 MSI time-series data (2020–2022). The analysis was conducted over 146 agricultural fields across multiple districts in Punjab and Haryana, representing heterogeneous smallholder conditions. Spectral analysis showed that SWIR1 (centered at 1613 nm) reflectance is sensitive to a small change in crop residue cover (CRC∼ 8-10%), exhibiting a strong linear relationship with CRC (R² = 0.97, p &lt; 0.01), while NIR/red-edge bands responded primarily to larger CRC changes (CRC &gt;20%). Temporal spectral profiles underscored pronounced distinctions in SWIR1, NIR, and red edge wavelengths corresponding to tillage intensity and CRC levels. Temporal analysis of spectral indices led to the development of a new index, CPSI (Conservation Practice Spectral Index) which integrates the features for sensitivity to non-photosynthetic vegetation, crop residue and tillage practice. Using the temporal maximum CPSI value (CPSI_max) during the sowing window and a single threshold derived from training data, the method achieved an overall classification accuracy of 92% with a Kappa coefficient of 0.81 on independent validation data, effectively distinguishing CT from HS/SS practices. District-wise accuracies ranged from 91% to 98%, demonstrating strong spatial and temporal robustness across years. This research delivers an efficient, scalable, and highly accurate methodology for early detection of conservation tillage, providing a critical tool for robust low-cost MRV in agricultural carbon credit programs and for the promotion of regenerative agricultural practices.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 101967"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147613421","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A metaheuristic design for hyperparameter tuning of XGBoost: Towards shaping an impeccable strategy for predictive modeling of mineral prospectivity","authors":"Saeid Esmaeiloghli ,&nbsp;Mahyar Yousefi ,&nbsp;Emmanuel John M. Carranza","doi":"10.1016/j.rsase.2026.101980","DOIUrl":"10.1016/j.rsase.2026.101980","url":null,"abstract":"<div><div>Over the past two decades, machine learning (ML) has been an avant-garde technology for data-driven mineral prospectivity mapping (MPM) due to its competence in modeling non-linear relationships embedded in multi-source geoscience datasets. XGBoost is a flagship algorithm in boosting-based ML that has gained popularity in MPM for its proficiency in speeding up the training process, enhancing prediction accuracy, reducing the risk of overfitting, and improving generalizability. However, achieving a well-trained XGBoost model usually entails careful tuning of multiple hyperparameters, i.e., <em>n_estimators</em>, <em>max_depth</em>, <em>min_child_weight</em>, <em>gamma</em>, <em>subsample</em>, <em>colsample_bytree</em>, and <em>learning_rate</em>. Motivated by this challenge, we conceptualized a computational framework in this paper that employs particle swarm optimization (PSO) to discover optimal XGBoost-related hyperparameters that yield MPM predictions with the highest accuracy. The PSO algorithm was engineered using a training dataset and a five-fold cross-validation strategy to find a hyperparameter setting that is globally optimal for achieving an XGBoost model with a robust performance. The PSO‒XGBoost model was coded and implemented by scripting over functions developed within the R programming language. The potential application of the proposed hybrid model was demonstrated through a real-case experiment for predictive modeling of porphyry-type copper prospectivity in the Baft-Sarduiyeh district, southern Iran. A comparative analysis between the PSO‒XGBoost model and manually tuned XGBoost scenarios revealed the superiority of the former by delivering higher values for confusion matrix-derived evaluation metrics and creating a higher-performance curve during receiver operating characteristic analysis. The findings suggest that tuning XGBoost with PSO-optimized hyperparameters can significantly improve predictive power and promote interpretability for more accurate modeling of mineral prospectivity.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 101980"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147613480","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Satellite-based soil texture mapping using GLOBAL–LOCAL spectral libraries for flood hazard assessment in arid regions","authors":"Abdel Rahman S. Alsaleh ,&nbsp;Mariam Alcibahy ,&nbsp;Delal E. Al Momani ,&nbsp;Hamed Al Hashemi ,&nbsp;Ali A. Al Hammadi ,&nbsp;Khaled AlAwadi ,&nbsp;Lakmal Seneviratne ,&nbsp;Maryam R. Al Shehhi","doi":"10.1016/j.rsase.2026.101997","DOIUrl":"10.1016/j.rsase.2026.101997","url":null,"abstract":"<div><div>Flooding in arid and semi-arid regions is a growing concern due to rapid urban expansion and more frequent extreme rainfall events, yet accurate flood hazard mapping remains difficult because of scarce in situ data and limited high-resolution soil information. To the best of our knowledge, this study represents one of the first demonstrations of satellite-based soil texture mapping via GLOBAL–LOCAL soil spectral library fusion for operational flood hazard assessment in an arid country. The fused model showed strong performance for sand (R² = 0.72, RMSE = 8%) and moderate performance for silt (R² = 0.44, RMSE = 2.8%) and clay (R² = 0.55, RMSE = 5.4%), while the Sentinel-2 land-cover classification used to mask non-soil surfaces achieved 93.75% overall accuracy (κ = 0.90). Combining predicted USDA soil texture classes with elevation, slope, hydrological, and land-cover data yielded a structural flood hazard map that highlights areas of higher susceptibility in low-elevation coastal zones and urban areas underlain by finer-textured soils, rather than predicting event-specific flood extents or water depths. Validation against Landsat 9 observations from a single major flood event in April 2024 showed strong spatial correspondence between high and very-high hazard classes and observed inundation in coastal cities and along major stream networks. By demonstrating the feasibility of transferring laboratory-trained soil spectral models to satellite sensors and embedding them in flood hazard mapping, this study establishes a scalable and transferable methodology applicable to arid and semi-arid regions worldwide where soil data scarcity limits hazard assessment. The resulting flood hazard and road vulnerability maps translate the scientific modelling framework into decision-ready products that support urban zoning regulations, drainage and transport infrastructure design and upgrading, and emergency response planning in rapidly developing arid and semi-arid regions.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 101997"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147657706","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Assessing spatial scale effects in multi-sensor fire fuel mapping in the heterogeneous landscapes of Tasmania","authors":"Sophia Hoyer ,&nbsp;Anke Fluhrer ,&nbsp;Florian Hellwig ,&nbsp;Steve Harwin ,&nbsp;Jukka Matthias Krisp ,&nbsp;Thomas Jagdhuber","doi":"10.1016/j.rsase.2026.101996","DOIUrl":"10.1016/j.rsase.2026.101996","url":null,"abstract":"<div><div>Effective fire-risk management in Tasmania requires vegetation maps that capture both broad fuel patterns and small, highly flammable gorse (<em>Ulex europaeus</em>) infestations. Yet gorse often occurs in fragmented patches that disappear in coarser land-cover products, raising uncertainty about how much fuel information is lost when regional maps are produced at moderate or coarse resolution. To clarify these scale effects we compare Object-Based Image Analysis-Random Forest fuel mapping at 0.5, 3 and 10<!--> <!-->m in a heterogeneous Tasmanian agricultural landscape using fused optical, LiDAR and SAR features. Beyond accuracy at each scale, we quantify how classes merge, disappear, or persist between resolutions using transfer matrices and analyse how large a gorse patch must be to remain detectable at coarser scales. F1 scores are consistently high across scales (76%–99%), yet class-level behaviour differs substantially. The 3<!--> <!-->m model achieves the highest gorse classification performance while maintaining geometric coherence of these shrub patches. When transferred from 0.5<!--> <!-->m, 76% of fine-scale gorse area remains represented at 3<!--> <!-->m, compared to only 36.8% at 10<!--> <!-->m. Detection probability at 3<!--> <!-->m increases monotonically with patch size, whereas at 10<!--> <!-->m even large patches (10,000–30,000<!--> <!-->m<span><math><msup><mrow></mrow><mrow><mn>2</mn></mrow></msup></math></span>) are detected in only 60% of the cases. These results demonstrate that high within-scale accuracy does not guarantee cross-scale persistence of fine-grained fuels. 3<!--> <!-->m resolution provides optimal scale–patch alignment for regional fuel-zone delineation in Tasmania, whereas sub-metre imagery is required for explicit identification of individual gorse infestations. Overall, the results confirm that spatial aggregation disproportionately affects narrow and fragmented vegetation types. Resolution choice is therefore not merely a technical setting, but a decisive factor in whether hazardous fine fuels remain visible in regional fuel assessments.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 101996"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147657708","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Evaluation of the ICEYE constellation for observing short-term surface displacements induced by volcanic activity","authors":"MinJeong Jo ,&nbsp;Stacey A. Huang ,&nbsp;Jeanne M. Sauber","doi":"10.1016/j.rsase.2026.102003","DOIUrl":"10.1016/j.rsase.2026.102003","url":null,"abstract":"<div><div>ICEYE operates the world’s largest SAR constellation, offering global imaging services with more than 60 X-band SAR satellites launched since its first mission in 2018. The continuous expansion of the ICEYE fleet enables short-repeat observations of the Earth, which is crucial for responding to natural disasters such as floods, earthquakes, and volcanic activities. While ICEYE data have demonstrated capabilities in amplitude-based analysis for Earth science applications, the potential for InSAR analysis as part of the international SAR constellation has only begun to be explored. In particular, global monitoring and rapid response capabilities for short-term volcanic activity benefits immensely from a dense, multitemporal satellite network. Here, we quantitatively assessed the InSAR capability and quality of ICEYE data for observing volcanic activity over two locations: Kı̄lauea volcano (Hawai’i, US) and Reykjanes Peninsula (Iceland). We examined important InSAR quality metrics, namely decorrelation characteristics and accuracy of deformation measurement relative to GPS/GNSS, and then analyzed the feasibility of measuring three-dimensional (3D) displacements. We found that ICEYE data quality was more variable and had higher error relative to GPS/GNSS than other X-band satellites we studied, in particular COSMO-SkyMed (CSK), TerraSAR-X (TSX), and TanDEM-X (TDX). Furthermore, as with many current NewSpace providers, data tasking can be challenging. Still, we were able to demonstrate initial 2D deformation retrieval, which can be expanded to 3D retrieval if multiple geometries can be leveraged. These capabilities can be combined with ICEYE’s very short temporal baselines to enable unprecedented rapid-repeat observations of volcanic activity. Despite the reduced InSAR performance of the ICEYE constellation relative to larger InSAR-capable satellites, the fleet contributes unique capabilities to the global SAR constellation for capturing deformation patterns associated with rapidly evolving volcanic events.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 102003"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147657709","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Enhanced ResNet for lake wetland components classification based on Sentinel-2 composites: A case study of Taihu Lake, eastern China","authors":"Yuxuan Hu,&nbsp;Jingyi Wang,&nbsp;Yatian Xu,&nbsp;Rui Qian,&nbsp;Mingshi Li","doi":"10.1016/j.rsase.2026.102012","DOIUrl":"10.1016/j.rsase.2026.102012","url":null,"abstract":"<div><div>Lake wetland ecosystems perform critical ecological functions such as water purification, biodiversity maintenance, and climate regulation, making accurate and fine lake wetland components classification essential for ecological health assessment and productivity accounting. Although deep convolutional neural networks (CNNs) have demonstrated strong potential in image recognition, their application to fine-grained classification of lake wetlands remains limited due to the complex spectral characteristics created by water, land and vegetation interactions. This study developed an enhanced Multi-level Dual-Attention(MLDA)-ResNet50 deep learning model using multi-temporal Sentinel-2 data to achieve integrated classification of wetland components in Taihu Lake. A median compositing strategy based on phenological windows was implemented to address severe cloud contamination. Sample scarcity issue at local scales was resolved through upsampling, enabling native-resolution pixel-level classification via probability-based sliding window accumulation. Key improvements on the architectural framework of CNN included: embedding CBAM modules within residual blocks to enhance the discriminative power for key feature extraction, Reducing spatial attention kernel to prevent edge distortion in upsampled data., and proposing a multi-level dual-attention feature fusion (MLDA-FF) mechanism to integrate shallow texture features with deep semantic features. Experimental results showed an overall accuracy at 95.6% in the classifications, representing a 4.8% improvement over the baseline ResNet50, with substantially escalated performance in spectrally similar land cover mixtures and small-scale feature areas. This research validates the applicability of CNNs for lake wetland components classification in Taihu Lake, and offers new methodological insights for future studies through its improved classification framework and data preprocessing strategy.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"42 ","pages":"Article 102012"},"PeriodicalIF":4.5,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147657813","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}