Issue Ebookcover storyIssue Contents
Latest Issue
Online FirstList of IssuesDocument Types
2026
Issue 9
Issue 8
Issue 7
Issue 6
Issue 5
Issue 4
Issue 3
Issue 2
Issue 1
2025
2024
2023
2022
2021
2020
2019
2018
2017
Abstract：Fractional Vegetation Cover (FVC) is a fundamental biophysical parameter that characterizes the structural properties of vegetation canopy and serves as a key indicator of terrestrial vegetation conditions. The FVC parameter plays an indispensable role in monitoring land surface processes, evaluating ecosystem health, and supporting studies on climate change and environmental dynamics. In recent years, continuous advancements in remote sensing technologies, particularly in spatial and temporal resolution, have substantially improved the development of FVC estimation algorithms. Therefore, quantitative remote sensing products derived from FVC data have been widely applied across various research domains, including climate variability analysis, ecological and environmental assessment, land use and land cover change monitoring, and carbon cycle modeling.However, most of currently available mainstream FVC products are generated at medium to low spatial resolutions, typically ranging from hundreds of meters to several kilometers. Despite the suitability of these products for large-scale and long-term vegetation monitoring, they exhibit notable limitations in terms of spatial detail and accuracy, particularly at local or regional scales. In areas with high spatial heterogeneity, mixed-pixel effects are prevalent, introducing considerable uncertainty into FVC estimations. These limitations pose remarkable challenges for precision ecological monitoring, refined land management, and informed environmental decision-making. Consequently, the demand for high-spatial-resolution (≤30 m) FVC products capable of supporting fine-scale vegetation monitoring and environmental assessment is increasing within the scientific community.This study aims to systematically review recent advancements in high-spatial resolution FVC products and remote sensing retrieval algorithms. First, a comprehensive summary of publicly available high-spatial-resolution (≤30 m) FVC products is presented, focusing on their spatiotemporal characteristics, accuracy, advantages, and limitations. Subsequently, four representative categories of FVC estimation approaches are reviewed: empirical regression models, spectral unmixing models, physically based models, and machine/deep learning models. For each category, their representative algorithms are highlighted, and their strengths and limitations are discussed. Finally, commonly used validation strategies and evaluation metrics for FVC products are examined, placing particular emphasis on key challenges, including the difficulty in acquiring reliable reference data, scale mismatches between reference data and target products, and the influence of spatial heterogeneity.Looking ahead, five key directions for future research are identified. First, the integration of big data and artificial intelligence, particularly deep learning frameworks and cloud-based computing infrastructures, is expected to optimize model generalization and improve retrieval efficiency. Second, the importance of leveraging high-resolution data from domestic Earth observation satellites, such as the GaoFen series, is emphasized, aiming to improve the spatial and temporal coverage of FVC products across diverse ecological zones. Third, the development of high-resolution FVC estimation models that combine data-driven approaches and physical mechanisms is highlighted as a promising direction for advancing quantitative remote sensing and vegetation cover inversion. Fourth, the establishment of standardized, high-fidelity reference datasets is crucial for robust product validation and inter-comparison, ensuring the consistency and interoperability of FVC datasets across scales and platforms. Finally, the feasibility of developing high-resolution FVC products at global or continental scales is evaluated. Overall, this study is expected to support the advancement of high-spatial-resolution FVC products and offer a theoretical reference for the optimization of retrieval algorithms and future product development.Keywords：fractional vegetation cover;High resolution satellite;quantitative remote sensing products;inversion;validation0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Data Assimilation (DA) integrates the process-based mechanisms of Crop Growth Models (CGMs) with the spatiotemporal continuity of Remote Sensing (RS) observations, enabling the large-scale and continuous monitoring of crop conditions and yield prediction. However, traditional assimilation frameworks have mainly relied on structural variables, which provide limited constraints on underlying physiological processes. With the rapid development of multisource RS observations and advanced computational methods, assimilable variables and algorithmic frameworks have evolved markedly. In this study, we aim to review systematically advances in RS DA for CGMs in recent years, with a focus on the expansion of assimilable variables, evolution of assimilation algorithms, and concept of assimilability. In addition, we propose an assimilability-based evaluation perspective to assess the suitability of CGMs in DA systems systematically.Our review is organized from two complementary perspectives: assimilated variables and assimilation algorithms. First, we analyze newly emerging RS variables, such as solar-induced chlorophyll fluorescence (SIF), vegetation optical depth (VOD), and synthetic Aperture Radar (SAR), and examine their physical mechanisms and potential for constraining crop physiological processes, canopy structure, and water dynamics. Second, we summarize recent advances in assimilation algorithms, including hybrid filtering methods, such as the particle filter–ensemble Kalman filter, and Machine Learning (ML)–enhanced frameworks, including surrogate models and physics-informed neural networks. These approaches not only enhance computational efficiency but also improve the robustness of covariance estimation in high-dimensional systems. Furthermore, we systematically evaluate the assimilability of mainstream CGMs, such as WOFOST, AquaCrop, DSSAT, and APSIM, from the perspectives of model structure, parameter identifiability, and observation operator construction.Our review identifies three key findingsEmerging RS variables significantly enhance the ability of DA systems to constrain key physiological and hydrological processes. SIF provides direct information on photosynthetic activity, VOD reflects vegetation water content and biomass dynamics, and SAR offers complementary observations sensitive to structural and moisture variations under all-weather conditions. However, these variables also introduce new challenges, including retrieval uncertainties, scale mismatches, and complex observation–state relationships. In terms of algorithms, hybrid filtering approaches improve the representation of nonlinearity and non-Gaussian distributions, whereas ML-based surrogate models further enhance computational efficiency and flexibility in observation operator construction. Assimilability analysis indicates that models with moderate complexity and clear observation–state mapping, such as WOFOST and AquaCrop, generally exhibit favorable performance in DA frameworks. In contrast,highly complex models, such as DSSAT and APSIM, are often constrained by parameter identifiability and weak observation–state coupling. Model structural bias and insufficient observation–state mapping are identified as key limiting factors affecting assimilation performance. Our findings collectively indicate a shift from structure-based constraints toward process-oriented assimilation in recent studies.The integration of advanced RS variables and innovative assimilation algorithms is driving a transition from traditional structure-based assimilation toward process-oriented constraints in CGMs. Future research should focus on improving observation operator design, reducing uncertainty propagation in RS-derived variables, and enhancing the integration of physical models with data-driven approaches. In addition, further efforts are needed to promote the integration of multisource and multiscale observations, enabling a comprehensive characterization of crop growth dynamics under varying environmental conditions. The concept of assimilability provides a useful framework for guiding model selection and system design in DA applications. Our review provides a comprehensive synthesis of recent developments and offers a structured perspective for future research in agricultural DA. These developments are expected to support the construction of next-generation agricultural monitoring systems with improved accuracy, robustness, and scalability, thereby contributing to sustainable agricultural management under changing environmental conditions.Keywords：crop growth models;remote sensing data assimilation;assimilability;solar-induced chlorophyll fluorescence;vegetation optical depth;synthetic aperture radar;hybrid filtering algorithms;machine learning-based assimilation frameworks0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Research Progress
Abstract：The micro laser retroreflector array INRRI (INstrument for Landing-Roving Laser Retroreflector Investigations) is one of the international payloads onboard the Chang’e-6 mission. This instrument receives laser signals emitted by the laser equipment onboard lunar orbiters and utilizes the parallel-reflection characteristics of incident laser on the reflective surface to achieve high-precision ranging. Through repeated observations and calculations, INRRI has become the first absolute control point established on the far side of the Moon, providing a solid foundation for lunar geodesy, remote sensing mapping and positioning, and high-precision orbit determination and navigation of lunar orbiters. To investigate the operating characteristics and application capability of micro retroreflector arrays in the lunar environment, this study presents the mechanical structure and optical design of INRRI and analyzes its key performance characteristics. In terms of design, the payload adopts a spherical dome structure and coating technology, enabling effective reception of the orbiter’s laser beam over a large field of view. In terms of performance analysis, an effective reflection area model is established to examine the reflected signal intensity under different incident angles. The results show that the effective reflection area of a single corner cube reflector reaches approximately 1 cm² under normal incidence, while the INRRI maintains stable reflection performance within a 60° half-aperture field of view. Furthermore, the detectability of INRRI by the laser equipment onboard lunar orbiters is comprehensively evaluated by considering the velocity aberration effect and far-field diffraction theory. Measurements obtained using a ZYGO interferometer indicate that the dihedral angles of the CCR are 0.37″, 0.64″, and 0.37″, corresponding to a total beam deviation angle of approximately 3.01″. Based on a far-field diffraction optical path and simulation analysis, the resulting far-field diffraction pattern exhibits spots with varying intensities distributed over an angular range of 40 μrad. This angular coverage exceeds the calculated maximum velocity aberration offset angle of 11.01 μrad, satisfying the requirements for orbital observation and compensation. The observability of INRRI has been further confirmed through multiple successful detections using the Lunar Orbiter Laser Altimeter onboard the Lunar Reconnaissance Orbiter. The above results demonstrate the feasibility of employing micro laser retroreflector arrays as absolute control points on the lunar surface and provide a reference for the design optimization and future application of micro laser retroreflectors in subsequent deep-space exploration missions.Keywords：Laser retroreflector;laser ranging;effective reflection area;velocity aberration;far-field diffraction512|142|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Dust aerosols play a crucial role in atmospheric radiative forcing, climate change, and significantly affect air quality. Therefore, accurate dust detection is essential for examining dust transport and supporting environmental monitoring. However, traditional dust detection methods, such as threshold-based physical indexes and conventional machine learning models, often struggle to efficiently capture the spectral and spatial features of dust, especially over heterogeneous land surfaces. Aiming to address this limitation, this study proposes an improved U-Net dust detection model that integrates a Spatial-Spectral Attention mechanism (SSA-UNet). The proposed model aims to enhance dust event detection using FY-4A/AGRI observations while providing a reliable and generalizable method for large-scale operational dust monitoring.The proposed SSA-UNet is based on the U-Net architecture with two key innovations. First, a dual-attention module combining channel and spatial attention is embedded into the encoder–decoder structure to adaptively emphasize informative spectral bands and crucial spatial regions associated with dust plumes. Second, residual skip connections are introduced to replace plain skip connections, mitigating gradient degradation in deep networks while improving training stability and feature reuse. For model inputs, a multidimensional “tile-scale” feature set is constructed, comprising multi-band reflectance and brightness temperature from the Advanced Geostationary Radiation Imager (AGRI) onboard FY-4A, dust-sensitive spectral indexes, and auxiliary meteorological parameters from ERA-5 reanalysis data (e.g., temperature, humidity, and wind fields). All input features are normalized and organized into spatial tiles that capture local details and the broader dust distribution context. The training dataset comprises manually labeled dust and non-dust samples collected from diverse seasons and underlying surface types. The model is trained using a loss function of binary cross-entropy.Quantitative evaluation on an independent validation set shows that SA-UNet achieves an overall accuracy of 99.46%, a mean intersection-over-union of 93.57%, and an F1-score of 87.17%. These results indicate that SA-UNet notably outperforms traditional physical index methods (e.g., brightness temperature difference approaches) and conventional machine learning models that rely only on spatial information. To further assess the model’s generalization capability, a separate validation experiment was conducted using an independent dust event dataset that was excluded from model training. In this challenge test, the dust identification results from SA-UNet were compared with manually interpreted data, yielding an overall accuracy of 98.77%. This high consistency confirms that the proposed model does not simply memorize training samples but efficiently learns the underlying spatial–spectral features of dust aerosols.Overall, this study demonstrates that integrating a spatial–spectral attention mechanism into a deep learning framework substantially improves the accuracy and robustness of satellite-based dust detection. The SA-UNet effectively overcomes the limitations of threshold-based indexes and spatial-only models by jointly leveraging the rich spectral information and spatial patterns of dust captured by the FY-4A/AGRI. The proposed method provides a reliable technical reference for operational, large-scale dust monitoring and dust transport analysis across China and its surrounding regions. Furthermore, this method shows a strong potential for extension to other geostationary satellite sensors and for aerosol type classification tasks.Keywords：dust detection;FY-4A satellite;U-Net;Spatial-Spectral Attention (SSA) mechanism;deep learning0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：The FY-3G satellite is the first active precipitation measurement satellite developed by China. The onboard Micro-Wave Radiation Imager-Rainfall Mission (MWRI-RM) has undergone a comprehensive upgrade compared with the previous-generation MWRI-I. Two additional oxygen detection bands and water vapor bands have been incorporated, allowing MWRI-RM to achieve integrated detections across 17 frequency points and 26 channels.The geolocation deviation of microwave remote sensing data directly affects the accuracy of subsequent multi-source product fusion and introduces non-removable errors. FY-3G is the first satellite in the Fengyun series to adopt a low-inclination orbit design. Aiming to ensure maximum energy capture by its solar panels, the satellite is designed an automatic steering strategy with an approximately 29.2-cycle to maintain a balanced energy supply. However, this strategy also leads to complex variations in the thermal environment, posing new challenges to the pointing accuracy and geolocation stability of various onboard instruments. In addition, the geolocation accuracy index has been improved from the MWRI-I 1 pixel (15 km) of the sub-satellite point to the 1 pixel (5 km) of the full field of view for MWRI-RM. Therefore, a rigorous observation geometry model must be constructed to address these challenges.First, the traditional observation geometry model of the conical-scanning microwave radiometer was reviewed. A precise vector-based observation geometry model was then derived based on the precise measurement coordinate system of the MWRI-RM instrument. Subsequently, a geolocation model was established to convert image coordinates into the instrument’s interior orientation elements. The on-orbit geolocation deviation characteristics of conical-scanning microwave images were analyzed, the conversion relationships among along-track deviation, across-track deviation, and scanning deviation were deduced, and two methods for characterizing geolocation deviations were proposed. The on-orbit geolocation deviations of MWRI-RM were estimated using the coastline edge-matching method, and an equivalent instrument mismatch angle correction model was further derived to transform image coordinates into the elements of the MWRI-RM interior orientation. Finally, the on-orbit data from FY-3G were used to evaluate and correct the geolocation deviations of MWRI-RM at different stages.Through the on-orbit tests in the first and second stages, the geolocation effect was verified through coastlines, proving that the proposed precise vector observation geometric model can observably reduce the initial on-orbit geolocation deviation compared to the traditional perspective model. Aiming to further improve the geolocation accuracy of the MWRI-RM, the deviation was corrected based on the coastline edge matching algorithm. The accuracy verification results from the second and third stages show that the correction model of the equivalent misalignment angle corrects the MWRI-RM geolocation accuracy to (-0.21±4.71) km (along-track) and (0.15±1.42) km (across-track). Following the on-orbit correction in the third stage, the periodic fluctuations in the along-track deviation coincided with the 29.2-day yaw maneuver cycle of the satellite platform. This observation indicates that changes in the thermal environment caused by the satellite turning have pronounced effects on the geolocation accuracy of the MWRI-RM.Results show that the precise vector geometric model substantially reduces the initial on-orbit geolocation deviation compared with the traditional viewing-angle model. Once on-orbit operations stabilize, the geolocation accuracy of MWRI-RM achieves the 5 km standard requirement and reaches an overall full-field accuracy of 1 pixel, thereby addressing the requirements for real-time operation and product generation.Keywords：FY-3 precipitation measurement satellite;micro-wave radiation imager-rainfall mission;conical scan;observation geometric model;precise vector model;on-orbit correction;geolocation deviation representation0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21Cloud detection of simulated geostationary microwave sounding satellite （FengYun-4/GeoMWS） over oceans AI Introduction
Abstract：Leveraging the advantage of high-frequency observations, geostationary satellites effectively overcome the limitations of polar-orbiting meteorological satellites in terms of long revisit cycles and insufficient timeliness for monitoring short-lived severe weather events. Compared with infrared bands, microwave radiation is less affected by clouds and exhibits excellent penetration capability through non-precipitating clouds. Consequently, the deployment of microwave sounding instruments on geostationary orbits has become a crucial development direction in meteorological observations. The Fengyun-4 Geostationary Microwave Sounder (GeoMWS), currently under development in China, aims to achieve all-weather, all-day, and high-frequency three-dimensional observations of cloud and precipitation systems, as well as their internal structures. This capability is important for monitoring rapidly evolving weather systems and improving Numerical Weather Prediction (NWP). However, a prerequisite for satellite data assimilation (whether under clear-sky or all-sky conditions) is the accurate identification of clear-sky and cloudy pixels, along with the assignment of appropriate observation errors. Existing cloud detection algorithms are predominantly designed for polar-orbiting satellites and are difficult to directly adapt to geostationary microwave observations. Therefore, this study aims to develop a specialized cloud detection method for GeoMWS to facilitate the effective assimilation of its data.This study introduces a fast and efficient cloud detection algorithm specifically tailored for the GeoMWS instrument. The proposed method constructs a cloud detection index using two window channels (Channels 2 and 4). This method fully accounts for the sensitivity differences of these channels across various latitudinal regions and introduces brightness temperature normalization to eliminate background noise attributed to temperature variations. Furthermore, the optimal threshold for the cloud detection index is scientifically determined by analyzing the evolution characteristics of the index under different classification criteria.Evaluation results demonstrate that the proposed method achieves robust cloud detection performance across different time periods. The Probability of Detection for clouds (POD_cld) and the Hit Rate both exceed 75%, indicating a strong capability for identifying cloudy scenes. Meanwhile, the False Alarm Rate for clouds (FAR_cld) is effectively controlled below 20%. Compared with previous methods applied in this context (which typically maintain a detection rate of approximately 60%), the proposed method exhibits considerable advantages in detection accuracy.This study successfully develops a cloud detection algorithm optimized for GeoMWS. The proposed method not only provides high-precision cloud detection information for data assimilation under clear-sky and cloudy conditions but also offers reliable theoretical and methodological support for future operational applications. In particular, for critical areas such as typhoon monitoring and NWP data assimilation, this method will substantially enhance the application effectiveness of geostationary microwave satellite data.Keywords：geostationary orbit;microwave sounding;geostationary microwave sounding;cloud detection over ocean;simulated data320|133|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Chinese Satellites
Abstract：Vegetation phenology reflects the integrated responses of ecosystems to climatic conditions, land surface modifications, and anthropogenic disturbances, serving as a critical indicator for assessing ecosystem responses to urbanization and climate change. Rapid urban expansion has substantially altered surface thermal environments, vegetation structure, and ecological processes, leading to differentiated phenological responses across urban cores, towns, and surrounding rural areas. Although previous studies have documented urban-rural phenological differences, most analyses have relied on moderate-resolution remote sensing products, which are sensitive to mixed-pixel effects and have limited capacity to resolve fine-scale urban heterogeneity. However, systematic investigations based on high-spatial-resolution data across large spatial extents remain insufficient. This study aims to (1) extract high-resolution vegetation phenology metrics for major Chinese cities using Sentinel-2 imagery, (2) quantify phenological differences along the urban-town-rural gradient, and (3) examine how these differences vary across vegetation types, climatic zones, and city-size categories.Using Sentinel-2 imagery from the Copernicus program (2019—2024), enhanced vegetation index (EVI) time series were constructed for 128 cities across China and their adjacent town and rural areas. A dynamic threshold approach was applied to extract the start of season (SOS) and end of season (EOS) from annual EVI trajectories. Aiming to ensure biological plausibility, phenological metrics were limited within reasonable day-of-year ranges. The derived phenological dates were then validated against ground-based observations and compared with the moderate-resolution MCD12Q2 phenology product to assess absolute accuracy using MAE and RMSE metrics. Urban-town-rural differences were also quantified for each city and subsequently analyzed across vegetation types, climatic zones, and city-size classes. Statistical comparisons and regression analyses were performed to evaluate spatial heterogeneity and scaling patterns.Validation results indicate that Sentinel-2-derived phenology exhibits substantially lower MAE and RMSE values than the MCD12Q2 product, demonstrating superior absolute accuracy in heterogeneous urban landscapes. Nationally, urban vegetation consistently demonstrates spring advancement and autumn delay relative to surrounding areas. On average, SOS in urban areas occurs 1.26 d and 1.49 d earlier than in towns and rural areas, respectively, whereas urban EOS is delayed by 1.51 d and 1.25 d, respectively, indicating an extended growing season in urban environments.Phenological responses substantially vary among vegetation types. Forest ecosystems exhibit the strongest sensitivity to urbanization, revealing the largest magnitude of spring advancement and autumn delay. In contrast, other vegetation types display comparatively moderate responses. Climatic background further modulates urbanization effects. The temperate climate zone shows the most pronounced urban-rural phenological contrasts, whereas subtropical and tropical zones exhibit weaker and less stable differences. City size also affects phenological patterns. The advancement of urban SOS generally intensifies with increasing city size, indicating a scaling effect associated with enhanced urban heat island intensity. However, the delay of EOS is more evident in small- and medium-sized cities and may weaken or even reverse in megacities, possibly due to complex interactions among thermal stress, vegetation management, and land surface heterogeneity.Overall, this study provides high-resolution, large-scale evidence of differentiated vegetation phenological responses to urbanization across China. The results reveal that urban expansion systematically modifies growing season dynamics, characterized by earlier spring onset and delayed autumn senescence. However, the magnitude and direction of these effects depend on vegetation type, climatic background, and city size. By integrating Sentinel-2 imagery with a dynamic threshold extraction framework, this study improves the quantitative reliability of urban phenology assessment compared with conventional moderate-resolution products. These findings enhance the understanding of how urbanization reshapes ecosystem seasonal dynamics and provide novel insights into the socio-ecological implications of phenological shifts. Overall, these results offer scientific support for sustainable urban planning, ecological infrastructure optimization, and climate adaptation strategies under continued urban expansion.Keywords：remote sensing;urbanization;vegetation phenology;sentinel-2;dynamic threshold method;gradient difference731|164|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Accurate detection of forest disturbances is crucial for evaluating forest ecosystem health, understanding carbon cycle dynamics, and maintaining biodiversity under accelerating global environmental change. However, with increasing anthropogenic pressures and climate variability, conventional disturbance detection approaches often suffer from temporal inconsistency, noise interference, and limited capability in distinguishing disturbance types. In particular, most existing methods tend to overlook the coupling between long-term vegetation trends and seasonal dynamics, leading to uncertainties in disturbance identification. Therefore, this study aims to develop a high-precision and robust forest disturbance analysis framework that enhances disturbance detection accuracy, allows reliable classification of disturbance types, and systematically reveals the spatiotemporal evolution patterns of forest disturbances across heterogeneous landscapes.Aiming to address these challenges, an integrated analytical framework that combines trend decomposition and periodic adjustment was constructed based on dense Landsat NDVI time-series data. The proposed framework decomposes vegetation signals into long-term trend components and seasonal periodic variations, effectively capturing gradual degradation and abrupt disturbance events. Furthermore, a multidimensional constraint mechanism in the time-frequency domain was introduced to mitigate temporal edge effects, reduce observational noise, and suppress pseudo-disturbance signals that commonly compromise time-series analyses. In addition, dynamic land-use transition characteristics before and after disturbance events were incorporated to identify disturbance types through the integration of spectral-temporal features and land-cover conversion pathways. This synergistic strategy improves the sensitivity of disturbance detection and the reliability of disturbance type classification.Results demonstrate that the proposed framework achieves a disturbance detection accuracy of 85.41% (Kappa = 0.85, RMSE = 3.81 a), substantially outperforming widely used datasets and algorithms. Specifically, the framework achieves an accuracy that is 18.93% higher than that of the global forest watch dataset (GFW) and outperforms the Landsat-based detection of trends in disturbance and recovery (LandTrendr) and continuous change detection and classification (CCDC) methods by 13.14% and 22.68%, respectively. Furthermore, the disturbance type classification accuracy reaches 86.64% (Kappa = 0.85), surpassing the annual China land cover dataset (CLCD) and CCDC by 6.07% and 26.22%, respectively. Spatial analysis indicates clear spatial differentiation in disturbance intensity, with higher intensity observed across the southwest and lower intensity in the northeast of the study area. Temporally, disturbance patterns exhibit alternating dominance between forest-to-non-forest and non-forest-to-forest transitions, reflecting the complex interactions among deforestation, regeneration, and land-use change processes.This study highlights a dynamic balance mechanism between human activity intensity and ecosystem resilience under spatial heterogeneity. The proposed multidimensional analytical framework effectively improves the precision, robustness, and interpretability of forest disturbance detection and classification, providing new insights into the underlying mechanisms of forest disturbance processes. Moreover, the generated high-resolution disturbance dataset offers a reliable temporal foundation for further applications, including regional forest carbon sink assessment, ecosystem stability evaluation, and sustainable forest management. Overall, this study advances methodological development in remote sensing-based forest monitoring and provides important scientific support for ecological conservation and land-use policy formulation.Keywords：forest disturbance monitoring;Landsat data;time series analysis;disturbance type classification;spatiotemporal patterns of forest disturbance in Yunnan Province0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21Remote sensing mapping of Xinjiang cotton based on the theory of stable classification with limited sample AI Introduction
Abstract：Xinjiang is the largest cotton-producing region in China, and accurately mapping its cotton distribution is of considerable importance for successful cotton production management. The accuracy of machine learning-based remote sensing mapping depends on the quantity and quality of samples. However, the extensive cotton cultivation area in Xinjiang increases sampling costs, making it difficult to address the demands of high-precision dynamic mapping.Based on the theory of stable classification with limited samples and considering the high spatial heterogeneity of Xinjiang, this study constructs sample sets and designs sampling strategies by incorporating geographic spatial constraints for different cotton-growing regions, including Northern Xinjiang, Southern Xinjiang, and Eastern Xinjiang. The study further explores the stability of cotton classification under limited sample conditions from two perspectives: minimum sample size and tolerance for erroneous samples.(1) In Northern Xinjiang, Southern Xinjiang, and Eastern Xinjiang, classification accuracy remained stable (with a reduction in accuracy of less than 1%) when the training samples accounted for 16%, 24%, and 14% of their corresponding sample sets, respectively. Similarly, stable classification performance was maintained when the proportions of erroneous samples were controlled within 25%, 25%, and 10%. (2) Cotton mapping in Xinjiang based on the minimum sample size achieved satisfactory mapping results, with overall accuracies of 90.56%, 83.17%, and 94.57% for Northern Xinjiang, Southern Xinjiang, and Eastern Xinjiang, respectively. At the county scale, the estimated cotton planting areas revealed a strong correlation with existing cotton distribution maps, yielding an R² of 0.98 and a root mean square error (RMSE) of 5384.42 hm2. (3) When the sample sets were transferred to 2018, a year when the interannual variation in cotton planting area remained below the identified tolerance for erroneous samples, the proposed method continued to yield satisfactory mapping results. The overall accuracies reached 95.37%, 88.85%, and 90.57% for Northern Xinjiang, Southern Xinjiang, and Eastern Xinjiang, respectively. At the county scale, the estimated cotton area shows a strong correlation with existing cotton distribution maps, with an R² of 0.98 and an RMSE of 6268.09 hm2.Aiming to address the long-standing reliance on massive quantities of high-quality training samples for cotton remote sensing mapping in Xinjiang, this study introduces the theory of stable classification with limited samples. Considering the pronounced spatial heterogeneity across Xinjiang, geographic spatial constraints were incorporated into sample set construction and sampling design across different cotton-growing regions. The stability of cotton classification under limited-sample conditions was investigated from two perspectives: minimum sample size and tolerance for erroneous samples. The main conclusions are as follows: (1) For the sample sets from Northern Xinjiang, Southern Xinjiang, and Eastern Xinjiang, a minimum sample size and a tolerance for erroneous samples were identified that ensure stable classification (ε<1%), which preliminarily validates the effectiveness of the theory of stable classification with limited sample for cotton classification in Xinjiang. (2) Cotton mapping in Xinjiang based on minimum sample size, as well as sample transfer mapping conducted within the identified tolerance for erroneous samples, both achieved satisfactory mapping performance, with both R² values reaching 0.98. These results demonstrate the robustness, stability, and interannual applicability of cotton mapping in Xinjiang based on the theory of stable classification with limited sample. This study provides a new theoretical basis and practical technical pathway for establishing a high-accuracy, high-efficiency remote sensing cotton mapping system in Xinjiang, a region characterized by complex planting structures and high sample acquisition costs.Keywords：Stable classification theory with limited sample;Xinjiang cotton;remote sensing mapping;machine learning;limited sample;geographic spatial constraints;sample transfer323|36|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Accurate monitoring of forest resources at the individual-tree level is essential for forest ecosystem management. Unmanned Aerial Vehicle (UAV) visible light (RGB) imagery provides a cost-effective, high-spatial-resolution data source for these wide-area monitoring tasks. Such imagery comprehensively records the fine contours of trees as well as the surrounding forest habitat. The use of panoptic segmentation technology allows for the unified interpretation and synchronous extraction of all forest elements. However, interpreting highly closed-canopy forest scenes remains a critical challenge. Traditional deep learning approaches often separate semantic segmentation of background elements from instance segmentation of individual trees, leading to severe pixel-level classification conflicts and inconsistencies in spatial topology. Furthermore, the limited spectral information in RGB imagery frequently causes severe spectral confusion among adjacent trees. Aiming to systematically address these challenges, this study proposes an end-to-end forest panoptic segmentation model, termed FSC-Mask2Former. The proposed FSC-Mask2Former is built upon the Mask2Former baseline and introduces two core architectural improvements specifically designed for the unstructured features of forests. First, a frequency-domain texture awareness module is incorporated into the feature extraction pathway to compensate for the loss of micro-texture details caused by spatial downsampling. Functioning as a learnable high-pass filter in the feature space, the module retains critical edge gradients. Second, an Instance-aware Query Contrastive (IQC) head is integrated into the output of the Transformer decoder to maximize the inter-class feature distance between spectrally similar tree species. By imposing an anisotropic constraint on the feature distribution, the IQC head expands decision boundaries and fundamentally suppresses category assignment conflicts. Aiming to evaluate the proposed model, a densely annotated dataset was constructed using UAV RGB imagery collected from Gaofeng Forest Farm in the Guangxi Zhuang Autonomous Region. Additional datasets from Genhe City in the Inner Mongolia Autonomous Region, Jixi County in Anhui Province, and Hengzhou City in the Guangxi Zhuang Autonomous Region were used to validate model transferability. Comprehensive experiments demonstrate that FSC-Mask2Former substantially outperforms existing mainstream networks. The model achieves an overall Panoptic Quality of 57.0%, revealing a substantial gain of 11.0 percentage points over the baseline. Most notably, the foreground Recognition Quality reaches 56.0%, representing a 12.0 percentage point increase. Visualizations confirm that FSC-Mask2Former effectively separates touching instances in high-canopy-closure forest areas, precisely outlines boundaries for morphologically irregular canopies, and maintains the spatial coherence of background elements. Furthermore, multi-region experiments indicate robust generalization capabilities across different geographical and ecological conditions. The proposed FSC-Mask2Former successfully addresses the bottlenecks of spectral homogeneity and task separation in UAV-based forest interpretation. The results indicate that accurate, full-element forest mapping can be realized using universally accessible UAV RGB imagery, providing a practical, robust, and highly cost-effective technical paradigm for modern forest resource monitoring.Keywords：UAV remote sensing;Panoptic segmentation;Mask2Former;Frequency domain analysis;contrastive learning;Individual tree recognition566|187|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Forestry and Agriculture
Abstract：Sea State level ranking is essential for ship navigation safety, marine risk assessment, and real-time route planning. While Synthetic Aperture Radar (SAR) can acquire wide-area ocean observations under all-weather and all-time conditions, direct sea state level ranking from SAR images remains challenging. Existing studies mainly estimate continuous wave parameters then convert them into levels, an approach that increases processing complexity and may limit response speed under rapidly changing sea conditions. Moreover, SAR images are affected by speckle noise, the large intraclass variability of sea-surface textures, and ambiguous transitions between adjacent levels. This study aims to develop an efficient ordinal SAR sea state level ranking method that improves feature robustness, represents texture distributions stably, and explicitly models the ordered structure of sea state levels.An ordinal sea state ranking framework based on adaptive spatial-frequency fusion and Kernel Mean Embedding (KME) distribution representation is proposed. First, an adaptive spatial-frequency fusion network (ASFF-Net) extracts discriminative features through parallel spatial-domain and frequency-domain branches. The spatial branch captures multiscale texture patterns, whereas the frequency branch enhances wave-related spectral structures. An adaptive gated fusion module then assigns content-dependent weights to the two branches to suppress speckle interference and strengthen effective wave information. Second, KME is introduced to model channel-wise feature distributions by mapping enhanced features into a reproducing kernel Hilbert space and approximating the distribution embedding with random Fourier features. This representation reduces the influence of intraclass texture variations. Third, a cumulative-link ordinal threshold prediction module decomposes the five-level ranking task into ordered binary decisions and outputs the final sea state level according to threshold crossing. A joint loss combining ordinal regression, unimodality regularization, and noise-perturbation consistency is used to optimize classification accuracy, rank consistency, and robustness to imaging noise. Experiments are conducted on two constructed SAR sea state datasets from Sentinel-1 and Gaofen-3, with labels defined in accordance with WMO sea state levels and verified by using significant wave height data from ERA5 and expert interpretation.Experimental results show that the proposed method outperforms traditional machine learning methods and representative deep learning baselines on both datasets. On Sentinel-1, the proposed model achieves an accuracy of 0.957, 1-OFF of 1.000, MAE of 0.097, and Macro-F1 of 0.975. On Gaofen-3, it achieves an accuracy of 0.950, 1-OFF of 1.000, MAE of 0.104, and Macro-F1 of 0.967. Compared with that of the best baseline, the overall accuracy of the model on Sentinel-1 and Gaofen-3 has improved by 2.4% and 2.3%, respectively. The cross-level misclassification rate has reduced to zero, and the 1-off accuracy reaches 100%. Ablation experiments further demonstrate that ASFF-Net improves stable feature extraction; KME reduces ordinal prediction offset; and the cumulative-link ordinal classifier, together with the three regularization terms, concentrates prediction errors within adjacent levels.The proposed adaptive spatial-frequency fusion and KME-based ordinal ranking method effectively addresses speckle noise, intraclass variability, and ordinal dependency in SAR sea state level classification. By integrating feature enhancement, distribution representation, and cumulative ordinal prediction, the proposed method improves the classification accuracy and physical rationality of the predicted levels. Results verify the effectiveness and generalization ability of the proposed method for different spaceborne SAR sensors, indicating its potential for rapid, reliable sea state monitoring and decision support in maritime applications.Keywords：synthetic aperture radar;sea state level;adaptive spatial-frequency fusion;kernel mean embedding;ordinal classification;speckle noise suppression0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Ebinur Lake, the largest saline lake in Xinjiang, is a critical ecological barrier in Northwest China. Secchi disk depth (SDD) is a key indicator of the lake’s optical characteristics, ecological health, and sensitivity to natural changes and anthropogenic activities. Aiming to overcome the constraints of traditional SDD monitoring methods, including high cost, limited spatial coverage, and operational challenges, this study utilized field-measured SDD data collected in 2023 alongside multisource satellite remote sensing imagery spanning from 2010 to 2024. The RG‒RR spectral feature (r = 0.81) was identified as the most sensitive indicator for SDD, and a remote sensing inversion index model (y = 7.174e18.629x) tailored for Ebinur Lake was developed. Validation results revealed an R² of 0.753, an RMSE of 3.7 cm, and a MAPE of 10.94%. The validated model was applied to examine the spatiotemporal patterns of SDD and its influencing mechanisms over the past 15 years. Results indicate the following: (1) Ebinur Lake exhibits generally low SDD (multi-year mean 13 cm), with significant positive correlation between interannual variation and water area (r = 0.59). Seasonally, SDD is highest in spring (17 cm), followed by autumn (12 cm) and summer (12 cm). Spatially, under the influence of northwest winds, SDD demonstrates a ring-shaped distribution pattern with higher and lower values in the lake center and along the shoreline, respectively. During wet years, the average transparency increases by 23.7% compared with dry years. (2) SDD variations are jointly driven by natural and anthropogenic factors. Daily wind speed exhibits an immediate negative correlation with SDD (r=-0.44), while meteorological factors (air temperature, evaporation, and precipitation) indirectly influence SDD through lake area regulation, with precipitation demonstrating a one-year lag effect. Additionally, agricultural expansion and increased impervious surfaces within the watershed exacerbate lake shrinkage via inflow reduction, further exerting an indirect negative effect on SDD. Under the combined influence of strong wind disturbances and human activities, SDD alone cannot directly indicate the trophic status of Ebinur Lake. Aiming to minimize misinterpretations of water quality due to the combined effects of natural and human disturbances, one recommendation is to appropriately reduce the weight of SDD in water quality evaluation systems to scientifically reflect actual aquatic environmental conditions. Overall, these findings provide theoretical support and practical references for aquatic environmental monitoring, ecological assessment, and sustainable management of lakes and other water bodies in arid regions.Keywords：Ebinur Lake;Secchi disk depth;multi-source remote sensing;spatiotemporal variation;influencing factors0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Nearshore underwater topography encompasses essential baseline information for the land-sea transition zone. Satellite-derived bathymetry based on ICESat-2 ATL03 photon point clouds continues to face several challenges, including high noise density, non-stationary along-track photon distributions, discontinuous extraction caused by sparse deep-water signals, and reduced refraction correction accuracy due to simplified treatment of sea-surface fluctuations. Aiming to address these issues, this paper introduces a hierarchical denoising method that integrates stepped along-track thresholding with MODWT(maximal overlap discrete wavelet transform)-based multiscale trend constraints. The proposed method first applies Gaussian fitting to determine the sea-surface elevation range, followed by an along-track stepped segmentation strategy to separate surface and subsurface photons. MODWT-based multiscale fitting is applied to surface photons to capture instantaneous sea-surface fluctuations, providing surface constraints for refraction correction. For subsurface photon processing, histogram compression combined with adaptive thresholding is used for coarse denoising. The low-frequency bathymetric trend extracted by MODWT serves as a prior constraint in an iterative dynamic-threshold filtering process. Aiming to optimize the retention of sparse deep-water signals, a photon densification strategy is introduced, followed by a sliding-window local statistical filter for final refinement, yielding continuous and stable bathymetric photon profiles. The proposed method was validated at Santa Cruz Island, Aldabra Atoll, and Saipan Island using NOAA airborne lidar bathymetry data for accuracy assessment. Results demonstrate that the proposed method achieves a maximum detectable depth of 50 m, with RMSE values of 0.340—0.711 m and correlation coefficients exceeding 0.99 when compared with airborne data. Compared with the ATL24 product and DBSCAN and OPTICS methods, the proposed approach exhibits superior robustness in noise suppression and signal continuity. Additionally, a Sentinel-2-based random forest bathymetric inversion experiment was conducted to demonstrate active-passive fusion, confirming the usability of the extracted underwater photons for bathymetric retrieval.Keywords：ICESat-2;MODWT;multispectral remote sensing;seabed elevation retrieval;nearshore bathymetry;data fusion0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Coastal wetlands, which serve as transitional zones between terrestrial and marine ecosystems, play vital roles in coastal hazard protection, biodiversity conservation, and climate regulation. Accurate identification of coastal wetland land-cover types is crucial for ecological assessment, protection, and restoration efforts. Hyperspectral images provide rich spectral information and show remarkable potential for fine-grained coastal wetland classification. However, the high heterogeneity of wetland environments, coupled with the costly and labor-intensive process of field investigation and manual annotation, often results in limited labeled samples. This limitation restricts the performance and generalization capability of conventional supervised models. Cross-Domain Few-Shot Learning (CDFSL) has emerged as a promising solution by transferring knowledge from labeled source domains to target scenes with scarce labels. However, existing CDFSL methods often lack prior knowledge guidance and suffer from high computational complexity during feature extraction.Aiming to address the above challenges, this study proposes a Dual-Prototype Cross-Domain Few-Shot Learning (DP-CDFSL) network for coastal wetland hyperspectral image classification. First, a State-Space Model (SSM)-based spatial-spectral feature extractor is designed to capture global spatial-spectral dependencies with linear computational complexity, thereby offering discriminative feature representation. Second, class-specific textual information is constructed by integrating the spatial distribution, spectral characteristics, and discriminative information of each land-cover category. A pretrained language model is then employed to generate class-level textual prototypes from these types of information. The generated textual prototypes are explicitly aligned with their corresponding visual prototypes, forming a dual-prototype framework guided by semantic priors. Furthermore, a sample-weighted domain alignment strategy is introduced to minimize the distribution discrepancy between the source and target domains. Therefore, this strategy encourages the network to focus on hard-to-align samples.Experiments were conducted on three UAV-based hyperspectral datasets acquired from the coastal wetlands of the Yellow River Delta to evaluate the effectiveness of the proposed method. Using only five labeled samples per class for training, the proposed DP-CDFSL achieved overall accuracies of (89.12±1.74)%, (88.21±0.36)% and (78.85±0.88)% on the three datasets. Compared with several state-of-the-art supervised and few-shot learning algorithms, DP-CDFSL consistently exhibits higher classification accuracy and produces more homogeneous classification maps with clearer land-cover boundaries. Ablation studies further confirm that each component, including the SSM-based feature extractor, the dual-prototype module, and the sample-weighted alignment strategy, contributes to the overall performance improvement.Overall, the proposed DP-CDFSL provides an effective solution for coastal wetland hyperspectral image classification under conditions involving limited labeled sample and domain shift. By aligning textual prototypes with visual prototypes, the proposed framework incorporates semantic priors into the representation learning process, thereby enhancing the discriminative capability of land-cover features. Meanwhile, the SSM-based spatial–spectral feature extractor efficiently captures global dependencies. In addition, the sample-weighted domain alignment strategy effectively improves cross-domain consistency by assigning considerable importance to hard-to-align samples.Keywords：coastal wetlands;hyperspectral images;cross-domain few-shot learning;prior knowledge;prototype network;state space model0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Atmosphere and Ocean
Abstract：Aiming to address the challenges of large-scale variations and feature extraction in small target detection for Synthetic Aperture Radar (SAR), this paper introduces an improved YOLO11 method with multiscale attention fusion.The proposed architecture replaces the original Bottleneck module with a multiscale feature-expanding residual module (MFDR) embedded within the C3K2 backbone network. The MFDR module applies hollow convolutions with four different hole rates to enhance multiscale feature extraction, while a contextual aggregation attention mechanism enables long-range feature interactions and extracts global contextual information. Additionally, a multiscale attention fusion enhancement module integrates channel and convolutional attention mechanisms for multi-granularity fusion, while three-branch convolution capture multiscale contextual information to improve feature representation for small and distant targets. The method further incorporates an Improved Concat_BiFPN module for dynamic cross-layer multiscale feature fusion.Experimental results demonstrate that the proposed approach achieves an accuracy of 98.7% on the SSDD dataset, with mean average precision (mAP@50) and recall rates reaching 99.2% and 95.3%, respectively, representing improvements of 4.6%, 2.1%, and 1.5% over baseline models. Conversely, on the HRSID dataset, the accuracy rate reached 93.1%, with mAP@50 and recall rates reaching 94.6% and 89.9%, respectively, corresponding to improvements of 0.3%, 4.5%, and 6.2%, compared with the baseline model.Overall, these results demonstrate that the proposed method effectively suppresses background interference and enhances the accuracy of SAR ship detection in complex scenarios.Keywords：SAR ship remote sensing image detection;attention mechanism;Small Target Detection;multi-scale feature extraction;deep learning;YOLO11n0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：With the rapid development of hyperspectral imaging technology, the high dimensionality of hyperspectral images poses considerable challenges for data storage and transmission. Resource-constrained scenarios, such as spaceborne platforms, have created an urgent demand for efficient onboard compression techniques. However, existing deep learning-based compressive sensing methods fail to realize an effective balance between reconstruction accuracy and computational efficiency under low sampling rates. In particular, Transformer-based models that rely on attention mechanisms incur computational and memory costs that grow quadratically with sequence length, leading to substantial efficiency bottlenecks when processing hyperspectral data. Aiming to address these issues, this study introduces the Mamba selective state-space mechanism, which offers linear computational complexity, to improve hyperspectral image reconstruction accuracy while maintaining low computational complexity. Additionally, a quantization-based coding strategy is further incorporated to enhance the overall compression performance of the model.A hyperspectral image compressive sensing network, termed Mamba-CSNet, is proposed by integrating the Mamba selective state-space mechanism with a two-stage quantization coding strategy. The network comprises three key components. First, a low-complexity dual-branch Mamba encoding structure is designed to capture long-range dependencies along the spectral and spatial dimensions. The resulting features are then integrated using cross-dimensional fusion to obtain joint feature representations. Second, an adaptive quantization and Brotli coding module is introduced to further exploit internal feature redundancy and reduce storage overhead. Finally, a CNN-Mamba feature enhancement module, termed CMM, is constructed to leverage the complementarity between CNN and Mamba. By collaboratively performing fine-grained local feature modeling and global context capture, the module effectively improves the representation and reconstruction capability of compressed features.Comparative experiments are conducted on a hyperspectral dataset constructed in this study, comprising 3016 samples, and Mamba-CSNet is systematically compared with five representative hyperspectral image compression methods. Experimental results demonstrate that, at a 1% sampling rate, the proposed method achieves the best balance between reconstruction accuracy and computational efficiency. Compared with the current state-of-the-art method, Mamba-CSNet improves the PSNR by 0.872 dB, reaching 38.024 dB, while reducing the computational cost by 11.1% to only 0.635 GFLOPs. Furthermore, owing to the proposed two-stage quantization coding strategy, the method exhibits excellent robustness in extremely bandwidth-limited scenarios. Even at an ultra-low bitrate of 0.042 bits per pixel per band (bpppb), the proposed method still achieves high-quality reconstruction with a PSNR of 37.058 dB.Through the integration of the Mamba selective state-space mechanism with a two-stage quantization coding strategy, Mamba-CSNet effectively enhances the modeling of cross-band correlations in hyperspectral images while addressing the limitations of existing learning-based models in terms of restricted reconstruction accuracy and high computational cost under low sampling rates. For practical applications involving sensors with varying spectral band configurations, future work will explore cross-band generalization mechanisms to achieve unified compression and high-quality reconstruction of hyperspectral images with different band configurations using a single general-purpose model.Keywords：hyperspectral image;compressive sensing;deep learning;Mamba selective state-space mechanism;quantization coding;low bitrate;spaceborne platforms109|27|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21A lightweight network for joint remote sensing image registration and change detection AI Introduction
Abstract：Image registration and change detection are two fundamental procedures for extracting and analyzing multitemporal information from remote sensing imagery. However, most deep learning-based approaches address these tasks separately. Existing joint frameworks either lack an explicit and interpretable registration mechanism or overlook the collaborative interaction between spatial and semantic difference features. This study aims to develop a lightweight, unified network that can align misregistered bi-temporal optical remote sensing images and accurately identify real land-cover changes under geometric offsets and pseudo-changes caused by spatio-spectral variations.A lightweight joint registration and change detection network, named LJRCDNet, is proposed. MobileNetV3-Large is first adopted as a shared encoder to extract four-scale features for registration and change detection. A spatial consistency module is then designed to perform semi-dense keypoint detection and local descriptor construction. Guided by self-supervised correspondences and knowledge distillation, the module estimates a homography-based spatial transformation to align multiscale feature maps of the pre-event image with those of the post-event image. In the aligned feature space, a spatiotemporal difference collaboration module is introduced. This module combines spatial cross-attention and channel cross-attention to model the coupling relationship between spatial and semantic difference features, while employing depthwise separable convolution and efficient channel attention to control computational cost. Finally, multiscale collaborative difference features are fused using a scale-adaptive perception module and upsampled to generate the final change map. Experiments are conducted on the SVCD, SYSU-CD, and SECOND datasets under co-registered and misregistered settings, and the proposed method is compared with 13 representative change detection networks.In the co-registered setting, the model without the registration branch still achieves competitive performance, with only 5.75 M parameters and 2.90 GFLOPs, which verifies the effectiveness of the difference collaboration design. In the misregistered setting, the complete LJRCDNet achieves IoU scores of 79.62%, 65.09%, and 52.24% on SVCD, SYSU-CD, and SECOND, respectively, outperforming the second-best methods by absolute gains of 4.52%, 5.56%, and 5.66%. Qualitative results show that the proposed method successfully suppresses false alarms caused by geometric offsets, illumination changes, seasonal variations, and differences in land-object appearance, while preserving notably complete boundaries of changed objects such as vehicles, impervious surfaces, cultivated land, and buildings. Ablation experiments further verify that SCM and STDCM are necessary: removing SCM reduces IoU by 1.87%, 3.03%, and 2.16%, while removing STDCM leads to reductions in IoU by 1.54%, 3.71%, and 1.29%, respectively. Replacing SCM with SIFT also degrades performance. Matching experiments further show that SCM achieves the best 3-pixel matching accuracy across all datasets.LJRCDNet integrates explicit spatial registration and spatiotemporal difference enhancement within an efficient end-to-end framework. The spatial consistency module provides reliably aligned features for subsequent change detection, while the spatiotemporal difference collaboration module enhances true spatio-spectral differences while suppressing pseudo-changes. The network achieves higher accuracy and better robustness compared with existing methods, while maintaining relatively low complexity, increasing its suitability for large-scale remote sensing image processing and deployment on resource-limited devices. Future work will extend the unified framework to heterogeneous remote sensing imagery and strengthen the robustness of the matching module in scenarios where large-area changes reduce the number of valid correspondences and degrade their spatial distribution.Keywords：remote sensing images;registration;change detection;Spatial consistency;spatio-temporal difference collaboration132|23|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：High-resolution remote sensing image change detection aims to distinguish real land cover or object changes from bitemporal observations and is widely used in urban expansion monitoring, land-use survey updating, disaster assessment, and ecological supervision. In high-resolution optical imagery, pseudochange responses are easily induced by illumination variation, cast shadows, seasonal phenology, radiometric shifts, and local registration errors. Meanwhile, multiscale aggregation and repeated upsampling may weaken the continuity of building boundaries, elongated structures, and small-scale changed regions, leading to fragmented predictions, internal holes, and locally over-smoothed outlines. These issues hinder lightweight models from simultaneously maintaining difference reliability, structural integrity, and computational efficiency.This study proposes TGS-Net, a high-resolution remote sensing image change detection network that combines reliable temporal difference screening with postfusion structural refinement, to address the abovementioned issues. The network is built on the dual-input encoder-decoder framework of ChangerEx and uses an interaction-enhanced MixVisionTransformer encoder to extract multiscale bitemporal features. The pipeline contains three stages: prefusion difference reliability screening, bitemporal alignment and fusion, and postfusion structure refinement. Before feature fusion, a Temporal Difference Gate (TDG) encodes absolute feature differences, generates lightweight gating weights, and symmetrically rewrites the two temporal branches. This design allows reliable change-related discrepancies to be preserved, whereas unstable appearance disturbances are suppressed before they propagate into subsequent fusion. After flow dual-alignment fusion, a large selective kernel (LSK)-based refinement branch is applied to the fused features. By combining efficient channel recalibration with large-kernel spatial context modeling, this branch enhances local boundary continuity and regional completeness and alleviates structural fragmentation in changed areas.Experiments are conducted on the WHU-CD, LEVIR-CD, and SYSU-CD datasets under consistent training and evaluation protocols. TGS-Net achieves F1 scores of 94.58%, 91.79%, and 81.11%, with corresponding intersection over union values of 89.72%, 84.83%, and 68.22%, respectively. Compared with the baseline model, the proposed network improves F1 by 0.62, 1.36, and 1.72 percentage points on the three datasets. Comparisons with recent methods show that TGS-Net maintains competitive accuracy with only limited complexity growth. Ablation experiments indicate that TDG and the LSK-based branch provide complementary gains. Feature-consistency statistics in unchanged regions and single-temporal spectral-shadow perturbation tests reveal that TDG reduces the sensitivity of the model to pseudochange interference in several challenging scenarios. Boundary F1 evaluation and local visual comparisons further demonstrate that the refinement branch improves boundary continuity and reduces fragmented predictions, especially in building-dominated scenes.Overall, compared with other methods, the proposed method establishes a clearer correspondence between two common error sources in high-resolution change detection and two targeted network operations: reliable difference screening before fusion and structural compensation after fusion. Results indicate that explicitly controlling unreliable temporal discrepancies then refining fused features with adaptive spatial context is an effective strategy for robust high-resolution remote sensing image change detection. At the same time, experiments reveal remaining limitations in complex mixed land-cover scenes, where weak low-contrast changes may still be missed and irregular boundaries may be locally over-smoothed. These observations provide a practical basis for future work on adaptive reliability estimation, boundary-preserving refinement, cross-scene generalization, and deployment-oriented acceleration.Keywords：high-resolution remote sensing;change detection;temporal difference reliability screening;temporal difference gate;pseudo-change suppression;large selective kernel;structural refinement0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Accurate road extraction from high-resolution remote sensing images is critical for applications such as urban planning, intelligent transportation systems, and emergency response. Despite recent advancements, state-of-the-art deep learning methods still struggle with road fragmentation, misclassification, and weak topological connectivity, especially in dense urban areas where roads are frequently occluded by buildings or shadows. A key limitation of existing methods lies in their main focus on road features, while neglecting the strong geometric and semantic priors provided by nearby buildings, which can serve as valuable contextual cues for inferring road continuity under occlusion. This study aims to develop a unified, instruction-driven framework that jointly leverages road and building semantics to enhance road extraction accuracy and network connectivity, even when building annotations are partially available.A novel instruction-guided network, Instruct-FHNet, is proposed for adaptive road-building perception. First, a mixed-instruction segmentation dataset is constructed by combining the Massachusetts and Paris road and building datasets, each paired with natural language instructions (e.g., “Please extract roads” or “Please extract buildings”). This design enables a single model to perform both tasks without requiring dual-label supervision for every image. The proposed architecture integrates four core components: (1) a visual encoder (ResNet34) and a text instruction encoder (Transformer-based); (2) a multilevel instruction semantic injection module that integrates task-specific textual guidance into visual features across multiple encoder stages; (3) a frequency feature and decoupling context capture module, which applies discrete wavelet transform to enable separation of low- and high-frequency components, followed by axial Transformers to model directional long-range dependencies; and (4) a Gaussian-convolution hybrid decoder that categorizes features into two branches—Gaussian smoothing for enhancing global connectivity and convolutional sharpening for improving local boundary precision—and then fuses them with skip connections for high-fidelity output.Experiments performed on the Massachusetts and Paris datasets demonstrate the state-of-the-art performance of Instruct-FHNet in road extraction. On the Massachusetts road dataset, the proposed framework attains an IoU of 68.31% and achieves the highest APLS, outperforming established models such as DLinkNet, UNetFormer, OARENet, and UGDNet. Ablation studies confirm that the frequency-decoupled context module substantially improves robustness to occlusions, while the instruction-guided semantic injection effectively leverages building priors to recover disconnected road segments, improving APLS by 0.52%. Notably, the model supports road and building segmentation through simple instruction switching after a single joint training phase. Although building extraction performance slightly decreases due to dataset imbalance, the framework maintains strong versatility. Visual results further validate its effectiveness in preserving road connectivity in complex scenarios, particularly in areas affected by building shadows and dense intersections.This study presents an innovative instruction-driven framework for road extraction that explicitly incorporates building semantic information as contextual guidance. Through the integration of multilevel instruction fusion, frequency-aware directional context modeling, and a hybrid decoding strategy, Instruct-FHNet effectively addresses key challenges of fragmentation and misclassification in occluded urban environments. The proposed framework’s effectiveness in performing dual tasks within a single model enhances practicality while reducing annotation dependency. This work not only advances road extraction accuracy and connectivity but also introduces new avenues for instruction-based, multitask learning in remote sensing image analysis.Keywords：remote sensing images;road extraction;building feature perception;semantic segmentation;instruct guidance0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Abstract：Accurate and efficient treatment of strongly forward-peaked scattering phase functions is a critical challenge in atmospheric vector radiative transfer, especially for remote-sensing simulations involving liquid water clouds, aerosols, and non-spherical ice crystals. The conventional δ-M truncation method can reduce the number of expansion terms required for radiative transfer calculations; however, it may still introduce systematic oscillations and contribute to accuracy degradation under extremely sharp scattering peaks. The δ-M+ method was developed as an extension of the δ-M method for scalar radiative transfer by introducing a smooth correction to high-order expansion coefficients. This study extends the δ-M+ method to vector radiative transfer calculations and evaluates its applicability, numerical accuracy, and computational efficiency in polarized radiative transfer simulations.The δ-M+ transformation was implemented in the vector radiative transfer model SOSVRT. Two vector extensions were considered: a matrix-form formula derivation method and a ratio-based correction method. The former derives the transformation coefficients for the independent elements of the scattering phase matrix based on the matrix representation of the truncated expansion, whereas the latter applies a ratio correction using the transformed scalar phase function. The accuracy of the transformed scattering matrix elements was assessed for representative scattering media, including aerosol particles, spherical liquid water cloud droplets, and non-spherical ice crystal particles. Full Stokes vector simulations were then performed to compare the performance of the δ-M+ methods with the conventional δ-M method under different optical thicknesses, viewing geometries, relative azimuth angles, scattering angles, and numbers of discrete ordinates.Results show that the δ-M+ formula derivation method effectively suppresses truncation-induced oscillations in the reconstructed scattering phase matrix while maintaining the polarization accuracy of vector radiative transfer calculations. For the scalar radiance component I, the δ-M+ method achieves comparable or superior accuracy to that of the conventional δ-M method while requiring fewer streams, indicating improvements in computational efficiency under strongly forward-scattering conditions. For the polarized Stokes components Q, U, and V, the δ-M+ formula derivation method generally maintains error levels comparable to those of the δ-M method under the same number of streams, indicating that a reasonable truncation treatment has minimal influence on polarization accuracy. In contrast, the δ-M+ ratio-based correction method shows limited applicability for media with relatively large optical thicknesses. The errors of this method become more pronounced for the U and V components and, under large scattering angles and greater optical thickness, may also increase the errors of the Q component and the cloud-bottom scalar radiance I.The δ-M+ method can be successfully extended from scalar radiative transfer to vector radiative transfer in a physically consistent and numerically stable manner. Among the two implementations examined in this study, the matrix-form formula derivation method provides better robustness and is more suitable for polarized radiative transfer simulations involving strongly forward-scattering particles. This method improves computational efficiency while maintaining the accuracy of Stokes vector calculations, provided that the truncation position is selected carefully based on the scattering properties and optical thickness of the medium. The proposed method also provides an efficient numerical treatment for vector radiative transfer simulations in satellite remote sensing, particularly for atmospheric scenes containing clouds, aerosols, and non-spherical ice particles.Keywords：polarization;the δ-M+ method;Legendre coefficients;vector radiative transfer;truncation of phase function;SOSVRT0|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-09-21
Models and Methods
