最新刊期
卷 30 ， 期 7 ， 2026
- 摘要：Remote sensing is a fundamental tool for understanding the Earth system. The integration of spaceborne, airborne, tower-based, and ground-based multi-platform collaborative observations with emerging technologies, including artificial intelligence, big data, and digital twins, is reshaping remote sensing paradigms. Traditional approaches based on static, single-sphere, and morphological observations are evolving toward a comprehensive framework that integrates dynamic, cross-sphere process inversion and 3-D perspective detection, thereby accelerating the transformation of geoscience toward a data-driven and quantitatively analytical paradigm.This study distills 10 cutting-edge scientific issues arising from the ongoing transformation of remote sensing and groups them into three major categories. The first category focuses on theory and modeling, covering multidimensional radiative transfer mechanisms, remote sensing foundation models and intelligent agents, and penetrating remote sensing of polar ice sheets. It aims to establish a unified modeling framework that integrates physical mechanisms with artificial intelligence. The second category centers on observational technologies, including virtual constellation collaborative observation, multimodal real-time data processing, as well as natural disaster monitoring and early warning, with an emphasis on advancing multi-platform collaboration and intelligent processing. The third category addresses systematic applications, involving carbon–water–energy cycle regulation, habitable planet exploration, human–environment system integration, and climate change impact assessment, aiming to meet critical scientific research and societal demands.These three categories collectively underline three key development directions: the in-depth integration of remote sensing and artificial intelligence, the synergistic utilization of multi-source geospatial data, and the systematic upgrade of the “observation-to-decision” paradigm. Specifically, theoretical research prioritizes breakthroughs in fundamental mechanisms, technological research focuses on systematic engineering implementation, and applied research targets interdisciplinary practical deployment. Overall, remote sensing science is transitioning from conventional experience-driven approaches to a novel paradigm that integrates physical mechanism constraints and data-driven analytics, while evolving from static environmental monitoring touard dynamic prediction and intelligent decision-making. Future research efforts should prioritize interdisciplinary collaboration, open scientific platforms development, and international cooperation to address major global challenges, including climate change, natural disaster outbreaks, ecological degradation, and deep-space exploration.关键词：radiative transfer;carbon-water-energy cycle;collaborative observation;remote sensing large model;AI intelligent processing;habitable planet;human-land system coupling;polar ice sheet;natural disaster;ecosystem124|20|0更新时间：2026-07-19
Scholar's View Point
-
摘要：Microwave Vegetation Optical Depth (VOD) is an important remote sensing parameter for characterizing vegetation water content, structure, and carbon-related ecosystem processes. As a result of the penetration and all-weather observation capabilities of microwave signals, VOD is less affected by signal saturation and cloud contamination than optical vegetation indices, providing unique information on vegetation water status and biomass dynamics. This review aims to summarize the theoretical basis, product development, validation methods, uncertainty sources, and major applications of VOD and identify key challenges and future directions for improving VOD retrieval and application.We systematically reviewed the development of VOD from three aspects: retrieval algorithms supporting product development, product validation and uncertainty assessment, and applications. First, the microwave radiative transfer theory and vegetation scattering mechanisms underlying VOD retrieval were summarized. Second, representative VOD retrieval frameworks and major products from active and passive microwave sensors were reviewed. Finally, existing validation strategies, uncertainty analyses, and application studies were synthesized to clarify the current status and limitations of VOD research.The review shows that the zeroth-order τ-ω model remains the dominant theoretical basis for passive microwave VOD retrieval, whereas inter-product consistency is largely constrained by the parameterization of essential radiative transfer variables, particularly effective scattering albedo and surface roughness. Multi-source fusion and advanced radiative transfer models provide promising pathways for improving retrieval accuracy and integrating multi-frequency VOD information. However, developing physically based and widely accepted parameterization schemes across frequencies remains a major challenge. Regarding product development, passive microwave observations have supported the development of multi-frequency, long-term VOD records, whereas long-term active microwave VOD datasets and global fusion products with broad spatial coverage and high spatial resolution remain limited. The absence of a globally consistent reference VOD dataset for product validation and uncertainty assessment indicates that current assessments still rely on indirect vegetation-related proxies, such as optical vegetation indices, aboveground biomass, and canopy height, complicating the differentiation of biomass-related signals from variations in vegetation water content. Current uncertainties in VOD products primarily arise from errors in observation and land surface temperature inputs; poorly defined retrievals; calibration issues of essential radiative transfer parameters; ancillary data errors; inconsistent frequency-dependent physical relationships; limited use of polarization information; and retrieval difficulties over complex surfaces, frozen ground, and snow-covered regions.Although current active and passive VOD products across different frequencies still require improvements in retrieval accuracy, inter-product consistency, and validation, existing studies have demonstrated that VOD, especially at low microwave frequencies, provides complementary and relatively independent information for constraining forest carbon stocks, detecting drought stress, and assessing fire risk. It also shows strong potential for land surface data assimilation and carbon–water cycle applications. Future research should focus on strengthening physical constraints regarding multi-frequency radiative transfer parameters; developing cross-mission and harmonized long-term VOD products with increased spatial resolution; improving validation systems via field observations, GNSS-based measurements, and triple collocation methods; and enhancing the synergistic use of VOD with soil moisture, other vegetation-related remote sensing indices, and land surface models. These advancements will improve the spatial and temporal resolution, physical consistency, and quantitative reliability of VOD-based vegetation monitoring and carbon cycle studies, providing a strong observational basis for ecosystem assessment and land surface process modeling under global change.关键词：passive microwave remote sensing;active microwave remote sensing;Vegetation Optical Depth (VOD);Vegetation Water Content (VWC);vegetation phenology applications;forest carbon sink monitoring38|14|0更新时间：2026-07-16 -
摘要：Understanding the mineralogical composition and abundance distribution of the lunar surface is essential for deciphering the origin and evolution of the Moon, assessing its resource potential, and supporting future crewed and robotic exploration missions. Owing to its wide spatial coverage, non-contact observation, and sensitivity to diagnostic spectral absorption features, optical remote sensing has become one of the most important techniques for lunar mineral detection. This review provides a systematic overview of optical remote sensing for lunar mineral exploration, focusing on its observational principles, payload development, data resources, inversion methodologies, representative achievements, current challenges, and future prospects.This review first outlines the physical basis of lunar mineral detection using optical remote sensing, including the spectral response mechanisms of typical lunar minerals and the influence of the lunar surface environment on reflectance spectra. It then summarizes representative optical payloads and data products from orbital and in-situ lunar exploration missions. Based on this, existing mineral identification and inversion methods are classified into three categories: spectral feature characterization with data-driven mapping, physics-based forward modeling with inverse retrieval, and spectral unmixing. Their theoretical assumptions, output variables, advantages, limitations, and applicable scenarios are compared to establish a unified methodological framework for lunar mineral remote sensing.Optical remote sensing has advanced lunar mineral detection from global mapping to regional interpretation and in-situ validation. Multispectral and hyperspectral datasets from Clementine, Kaguya, Chang’E-1, Chandrayaan-1, and other missions have enabled the mapping of FeO, TiO₂, optical maturity, and major minerals such as pyroxene, olivine, plagioclase, spinel, and volcanic glass. These observations have revealed compositional differences between mare basalts and highland anorthosites, as well as possible exposures of deep-seated materials around impact basins and crater margins. In-situ spectral measurements from Chang’E-3, Chang’E-4, Chang’E-5, and Chang’E-6 have further provided local constraints and validation for orbital retrievals. In addition, the detection of OH/H₂O absorption near 3 μm has expanded lunar optical remote sensing from mineral mapping to hydration detection and resource-related studies.Optical remote sensing has become a key approach for linking lunar mineral composition, geological evolution, and resource assessment. However, lunar mineral retrieval still faces uncertainties caused by space weathering, grain-size variation, illumination and shadow effects, preprocessing errors, nonlinear spectral mixing, and limited ground truth validation. Future studies should focus on physically constrained preprocessing, nonlinear unmixing, uncertainty quantification, and the fusion of orbital, in-situ, laboratory, and multimodal remote sensing data. These efforts will improve the accuracy, reliability, and geological interpretability of lunar mineral detection and provide stronger support for future lunar exploration, resource evaluation, and in-situ resource utilization.关键词：lunar;optical remote sensing;in-situ exploration;orbital remote sensing;mineral identification;spectral unmixing;abundance estimation13|13|0更新时间：2026-07-16 -
摘要：Medicinal plants are the material basis for the inheritance and development of traditional Chinese medicine. They contain a variety of bioactive components and have a wide range of pharmacological effects. The quality of medicinal materials varies greatly under different geographical and climatic conditions. At present, medicinal plants have shortcomings in delicacy management, sustainable utilization of resources, dynamic monitoring of pests and diseases, and growth warning. Traditional research methods are time consuming, laborious, and destructive sampling. As a nondestructive monitoring method, remote sensing technology is capable of rapidly and systematically acquiring large-scale, quantitative, and periodic information, which has been widely used in the production process of Chinese medicinal materials. On the basis of the development process of remote sensing technology and the current research status of medicinal plant resources, this study systematically reviews the application progress of remote sensing technology in four aspects, namely, resource inventory and investigation, quality assessment and analysis, growth status monitoring, and stress detection and monitoring of medicinal plants. By meticulously examining these application domains, this review elucidates the distinctive advantages and transformative potential of remote sensing technology in overcoming the inherent limitations of conventional monitoring approaches. Currently, remote sensing technology has evolved into a comprehensive perception system characterized by the complementary integration of satellite-based, unmanned aerial vehicle, and ground-based remote sensing data. Meanwhile, this technological framework continues to expand its monitoring dimensions, enhance its detection depth, and diversify its application scenarios. Consequently, remote sensing demonstrates substantial potential and broad prospects in the domain of medicinal plant research, offering unprecedented opportunities for large-scale, multitemporal, and high-precision investigation of medicinal plant resources, their spatial distribution patterns, growth dynamics, and ecological environments. This review systematically summarizes and critically evaluates the existing limitations and deficiencies inherent in current research endeavors. Building upon this comprehensive analysis, we propose several strategic directions for future advancement, including the establishment of a collaborative sensing system based on multisource active remote sensing, the in-depth exploration of underlying biophysical and biochemical mechanisms, the implementation of full-cycle and cross-scale monitoring, and the promotion of interdisciplinary integration. These proposed orientations aim to provide innovative theoretical frameworks and practical approaches for leveraging modern scientific and technological advancements to drive the sustainable development of the traditional Chinese medicinal material industry, thereby contributing to the preservation, rational utilization, and industrial upgrading of valuable medicinal plant resources. In summary, this study systematically reviews the research progress of remote sensing technology in the field of medicinal plant resources from the perspectives of resource investigation, quality analysis, growth status, and stress monitoring and comprehensively summarizes the relevant achievements and existing challenges. It aims to provide a reference for related research worldwide and promote the further in-depth application of remote sensing technology in medicinal plant resource studies.关键词：medicinal plants;remote sensing technology;monitoring;resource survey;quality analysis;environmental stress8|11|0更新时间：2026-07-16
Reviews
-
摘要：Water surface sun glint is a significant source of contamination in water remote sensing imagery. Sun glint phenomena are widespread in Sentinel-2 MSI images, severely limiting the application of Sentinel-2 MSI data in water remote sensing. From an empirical perspective, this study conducted sun glint distribution extraction, analysis of spatial-temporal patterns, and investigation into the causes of sun glint in five typical study areas. This study developed a water surface sun glint extraction algorithm based on the UNet model, with an overall accuracy exceeding 96%. Research on the spatiotemporal distribution patterns of water surface sun glint was conducted on the basis of the extraction results. Water surface sun glint in Sentinel-2 MSI imagery exhibits distinct regional and temporal distribution patterns. In some severely affected areas, the frequency of sun glint occurrence reaches up to 80%, with the average glint-affected area exceeding 30% per image. The occurrence of water surface sun glint is primarily influenced by solar angle, viewing angle, latitude, and orbital position, while inland small water bodies show minimal direct correlation with meteorological factors such as wind speed and air temperature. Adjusting imaging time, viewing angles, and imaging methods can effectively reduce the frequency and intensity of sun glint. On the basis of the quantitative evaluation and empirical analysis of water surface sun glint in Sentinel-2 MSI imagery, this study validates the formation mechanism of sun glint, providing data support for physical models and the mitigation of water surface sun glint in subsequent satellite missions.关键词：Sentinel-2 MSI;sun glint;spatial-temporal distribution;formation factors10|13|0更新时间：2026-07-16 -
摘要：To evaluate the ecological impacts of the “Purse Seine Demolition of Lakes and Reservoirs” policy in the Taihu Basin, this study investigated the spatiotemporal response of aquatic vegetation in four typical enclosure-aquaculture lakes: East Taihu Lake, Gehu Lake, Changdang Lake, and Yangcheng Lake. The study aimed to clarify how the expansion and subsequent removal of enclosure aquaculture affected submerged aquatic vegetation, floating-leaved/emergent aquatic vegetation, and algal bloom dynamics, and to provide scientific support for future lake ecological restoration and long-term management.Long-term Landsat imagery (1985—2024) and Sentinel-1 synthetic aperture radar (SAR) data (2016—2024) were used to reconstruct the spatiotemporal dynamics of enclosure aquaculture and aquatic vegetation. A U-Net model was applied to extract enclosure aquaculture areas from SAR imagery, while a vegetation-algal bloom extraction algorithm based on Landsat data was used to classify submerged aquatic vegetation, floating-leaved/emergent aquatic vegetation, algal blooms, and open water. Time-series analysis, spatial overlay, and correlation analysis were then conducted to compare vegetation changes before and after net-pen removal. The responses of aquatic vegetation were further examined across net-retained, net-demolished, and net-free areas to identify spatial heterogeneity and possible ecological mechanisms.The total area of enclosure aquaculture in the four lakes increased rapidly after the 1980s, peaked at 361.9 km² in 2001 and then declined continuously. By 2024, only 8.7 km² of enclosure aquaculture remained, indicating the substantial effectiveness of the policy in removing net pens. Aquatic vegetation showed a clear long-term transition characterized by a decline in submerged aquatic vegetation, a gradual increase in floating-leaved/emergent aquatic vegetation, and an overall reduction in total vegetation area. After 2017, submerged aquatic vegetation decreased sharply from 63.6 km² to 10.9 km², representing an 82.9% reduction. By contrast, floating-leaved/emergent aquatic vegetation became more dominant in several lakes, although its coverage also fluctuated in later years. Vegetation degradation was especially severe in net-demolished areas, where submerged vegetation often declined substantially or disappeared, and algal blooms occurred simultaneously. Net-retained areas generally maintained higher vegetation coverage and more stable conditions, suggesting that enclosure structures and associated artificial maintenance previously helped sustain submerged vegetation through physical support and management.The “Purse Seine Demolition of Lakes and Reservoirs” policy effectively reduced enclosure aquaculture and contributed to water-quality improvement in the Taihu Basin. However, results show that removing net pens alone did not guarantee aquatic vegetation recovery. Under long-term eutrophication, the combined effects of loss of physical support, reduced artificial maintenance, altered hydrodynamic conditions, and persistent internal nutrient loading may have accelerated submerged vegetation degradation. Future restoration efforts should therefore shift from single water-quality improvement measures to integrated ecological restoration, including internal nutrient reduction, submerged vegetation recovery, hydrological regulation, near-natural restoration, mapping of historical vegetation distributions, and long-term remote-sensing monitoring, to support adaptive lake management.关键词：Taihu Basin;enclosure aquaculture;Aquatic vegetation;alternative stable states;remote sensing;Purse Seine Demolition of Lakes and Reservoirs;ecosystem recovery;Yangtze River’s Great Protection Strategy7|12|0更新时间：2026-07-16 -
摘要：The Yellow River is characterized by hyperconcentrated sediment loads and highly imbalanced water–sediment relationships, making accurate retrieval of Suspended Sediment Concentration (SSC) essential for water–sediment regulation, flood control, and ecological protection. However, remote sensing retrieval of SSC in hyperconcentrated waters remains challenging because of spectral saturation in the red and near-infrared bands, attenuation in shortwave infrared bands, and complex nonlinear relationships between optical signals and sediment properties.To address these challenges, this study developed a physics-constrained multitask Convolutional Neural Network (CNN) framework for SSC retrieval and investigated the spatiotemporal dynamics of SSC along the Yellow River mainstem during 2013—2023. Based on daily SSC observations from 12 hydrological stations and corresponding Landsat 8 OLI imagery, 676 high-quality matched samples were constructed. A novel spectral index, termed the red and near-infrared to all shortwave infrared normalized ratio (RANS), was proposed to characterize the relative relationship between scattering enhancement in the red/near-infrared bands and absorption suppression in the shortwave infrared bands under high-turbidity conditions. RANS showed a strong correlation with ln(SSC) (Pearson r=0.901, R2=0.81). To incorporate this physically meaningful spectral information, RANS was introduced as an auxiliary regression task within a multitask CNN framework. Bayesian optimization based on a Gaussian process surrogate model was further employed to automatically optimize model hyperparameters.Results demonstrated that the proposed multitask self-optimizing CNN outperformed both the single-task CNN and the non-optimized multitask CNN. In the ln(SSC) domain, the model achieved R2=0.843 and RMSE=0.614, while in the original SSC domain it achieved R2=0.837 and RMSE=1256.75 mg/L, substantially improving upon the single-task CNN. Five-fold cross-validation confirmed stable generalization capability (R2=0.655±0.059, RMSE=1620±192 mg/L). The highest retrieval accuracy was observed within the 2000—5000 mg/L concentration range (R20.81, MAPE23%). Monte Carlo experiments further demonstrated superior model robustness and reduced uncertainty.The reconstructed SSC dataset revealed pronounced spatiotemporal heterogeneity along the Yellow River mainstem. Spatially, SSC exhibited a distinct pattern of low values upstream, peak concentrations in the middle reaches, and declining concentrations downstream. The Longmen-Tongguan reach showed the strongest sediment accumulation, whereas the Xiaolangdi-Huayuankou reach displayed significant sediment reduction during flood periods. Seasonally, SSC consistently remained low in spring, peaked in summer, and gradually declined in autumn.These results indicate that integrating the RANS spectral index with multitask learning and Bayesian optimization effectively improves SSC retrieval in hyperconcentrated waters. The proposed framework provides both a reliable methodology and long-term dataset support for large-scale sediment monitoring, hydrological modeling, and water–sediment management in sediment-laden river systems.关键词：the middle reaches of the Yellow River;hyperconcentrated sediment-laden water;suspended sediment concentration retrieval;Convolutional Neural Network (CNN);Multi-task learning;Bayesian optimization;Landsat 8 multispectral imagery8|10|0更新时间：2026-07-16 -
摘要：The riparian zone serves as a critical ecotone between terrestrial and aquatic ecosystems, playing a vital role in maintaining riverine ecological security and regulating water—land interactions. Accurate delineation and dynamic monitoring of riparian zones are fundamental for watershed-scale ecological protection and spatial management.This study developed a novel remote sensing method for large-scale riparian zone delineation based on the inundation frequency index, using multi-temporal Sentinel-2 imagery and the GEE (Google Earth Engine) platform. By constructing a water inundation frequency index and implementing a threshold-based partitioning approach, we objectively defined the inner and outer boundaries of the water-level fluctuation zone in the middle–lower reaches of the Yangtze River.(1) The delineated riparian zone has an average width of 608 m, ranging from 289 m to 982 m across different urban sections; (2) the ecological riparian ratio is 45.9%, exhibiting distinct spatial patterns of “middle reaches lower reaches,” “right bank left bank,” and “water-level fluctuation zone terrestrial buffer zone;” (3) the development utilization rate of riparian zones is 54.1%, primarily occupied by agricultural land (33.7%) and industrial-mining land (7.2%); and (4) approximately 87 km of hardened embankments are concentrated in the lower reaches, substantially disrupting natural land-water interactions. Furthermore, we observed a net reduction of 6.6 km² in ecological land area during 2022—2023, indicating a continuing trend of “ecological retreat and artificial expansion.”This study confirms that the inundation frequency-based remote sensing method enables continuous, automated, and large-scale delineation of riparian zones, providing robust technical support for watershed-scale ecological assessment and spatial management strategies.关键词：inundation frequency;ecological space;delimitation of riparian zone;remote sensing;middle-lower reaches of Yangtze River10|12|0更新时间：2026-07-16 -
摘要：Shorelines of rivers and lakes constitute the critical interface between inland water bodies and adjacent terrestrial systems, functioning as key transitional zones that regulate flood storage, buffer hydrological extremes, and maintain aquatic–terrestrial ecological systems. Accurate and timely monitoring of the dynamics of shorelines of rivers and lakes is therefore of great practical significance for flood risk mitigation, aquatic ecological protection, and ecological restoration. To address the insufficient accuracy of regional-scale identification of water extents and shorelines of rivers and lakes, as well as the limited understanding of their long-term dynamic changes, this study takes Chenzhou City as the study area and aims to develop a regional identification method for water extents and shorelines of rivers and lakes based on the Swin-UNet architecture, using Gaofen and Sentinel-2 remote sensing imagery.Deep learning models based on the Swin-UNet architecture were established to identify water extents, the boundaries of which were then vectorized to extract shorelines of rivers and lakes in the study area using Gaofen and Sentinel-2 remote sensing imagery. Based on the trained and validated Swin-UNet models, shorelines of rivers and lakes in Chenzhou City from 2015 to 2025 were extracted, and their long-term dynamic characteristics were analyzed in terms of shoreline length, morphological stability, and shoreline composition.The Swin-UNet models achieved a water body identification accuracy exceeding 90%. The multi-year average area of water extent in Chenzhou City was approximately 314.7 km2, with a corresponding length of shorelines of rivers and lakes of about 7,306.4 km, and the results derived from Gaofen imagery were highly consistent with those from Sentinel-2 imagery. The length of shorelines of rivers and lakes in Chenzhou City showed stronger intra-annual variability, with a coefficient of variation of 12.8%, than interannual variability, with a coefficient of variation of 9.0%. Shorelines of artificial water bodies were more morphologically stable than those of natural water bodies. Over the past decade, the composition of shorelines of rivers and lakes in the study area exhibited a transition from semi-natural/semi-artificial shorelines to artificial shorelines, while natural shorelines remained relatively stable.The identification method for water extents and shorelines of rivers and lakes based on the Swin-UNet architecture can effectively improve the accuracy and stability of automatic shoreline extraction in complex regions and provide technical support for long-term monitoring of regional shoreline dynamics. The findings of this study can provide a reference for the resource management of shorelines of rivers and lakes, aquatic ecological protection, and territorial ecological restoration in Chenzhou City.关键词：river and lake shorelines;dynamic monitoring;Swin-UNet;Gaofen satellite;Sentinel-2;Chenzhou City568|747|0更新时间：2026-07-16 -
摘要：River discharge is a pivotal variable within the hydrological cycle, holding significant importance for flood warning, water resource allocation, and eco-environmental management. Traditional ground-based methods are limited by the sparse distribution of stations and the high cost of data acquisition, particularly in areas with complex terrains or remote regions, making it difficult to meet the demands of precise water resource management. Satellite remote sensing technology offers extensive coverage and high spatiotemporal resolution, providing new data sources and methodologies for river discharge monitoring. Machine Learning (ML) approaches can accurately simulate complex relationships between river discharge and multiple driving factors, offering novel avenues for processing intricate hydrological data and optimizing models. Integrating ML algorithms with remote sensing and in-situ river discharge can provide an innovative measure for efficient and reliable river discharge monitoring.This study selected the Tangnaihai Hydrometry Station as the research area and proposed a river discharge monitoring method by integrating satellite remote sensing and ML methods. Sentinel-2 imagery was utilized to extract river water surface width on the basis of the Google Earth Engine cloud platform. Five GLDAS v2.2 model-simulated variables, namely, evapotranspiration, soil moisture, temperature, terrestrial water storage, and runoff, served as predictor variables. Discharge monitoring models were subsequently developed on the basis of four statistical methods (linear, power, exponential, and polynomial functions) and four ML algorithms (XGBoost, random forest, LightGBM, and CatBoost). The discrepancies among different models were assessed, and the shapley additive explanation (SHAP) method was employed to quantify the importance of different input variables.Results indicated that the polynomial function model demonstrated superior performance over the other three statistical models during the testing period, with a coefficient of determination (R2) of 0.67, and its error metrics (mean absolute error [MAE]: 319.01 m³/s, root-mean-square error [RMSE]: 393.14 m³/s) were lower than those of the other three models. Compared with the traditional statistical approaches, the ML models exhibited significant improvements overall in simulation accuracy and stability; the R2 increased by 46.15%, while the RMSE and MAE decreased by 54.61% and 55.65, respectively. Notably, the random forest model achieved the optimal performance in the testing phase, with an R2 of 0.96, a Nash-Sutcliffe efficiency coefficient of 0.89, an RMSE of 172.81 m3/s, and a MAE of 147.33 m3/s, reflecting robust generalization capability and stability. SHAP analysis revealed that water surface width contributed most significantly to the discharge monitoring model (189.02), followed by soil moisture (145.11) and temperature (97.41). The runoff variable exhibited the minimum degree of influence on the river discharge monitoring model with a value of 14.14%.This study confirms the feasibility and superiority of integrating satellite remote sensing and ML approaches for high-accuracy discharge estimation in regions characterized by complex topography and data scarcity. Future work could be optimized by integrating higher-resolution satellite imagery and mechanistic models with physical processes.关键词：satellite remote sensing;machine learning;statistical models;SHAP method;discharge monitoring;Tangnaihai Hydrometry Station380|866|0更新时间：2026-07-16
Remote Sensing Monitoring of Aquatic Eco-environment
-
摘要：Vegetation phenology serves as a critical indicator for assessing the ecological impacts of climate change. While most previous studies have concentrated on the interannual variations of specific phenological events, the intraseasonal asymmetry within the growing season remains poorly understood. This study aims to fill this knowledge gap by quantifying the asymmetry between the green-up phase and the senescence phase from the dual perspectives of temporal duration and productivity levels, represented by mean NDVI. Focusing on the temperate regions of China, this research seeks to reveal the spatial–temporal patterns of phenological asymmetry and clarify the underlying climatic regulation mechanisms driven by temperature and precipitation imbalances. By examining these dynamics over 41 years, this study provides a nuanced understanding of how ecosystems reorganize their seasonal cycles in response to global warming. For this assessment, long-term GIMMS NDVI 3G+ data covering the period from 1982 to 2022 and NOAA climate records were utilized. This study constructed two primary indices: the vegetation phenology asymmetry index (VPAI), which includes length-based asymmetry (VPAIL) and NDVI-based asymmetry (VPAIV), and the Climate Asymmetry Index (CAI), accounting for temperature (CAITemp) and precipitation (CAIPrcp). The growing season dynamics were analyzed north of 30° N to detect trends in the start of growing season (SOS), end of growing season (EOS), and length of growing season (LOS). Furthermore, a self-organizing map (SOM) classification was implemented to examine the spatial consistency between phenological parameter distributions and vegetation types, evaluating the potential of these metrics for large-scale vegetation mapping and identifying the distinct signatures of croplands in comparison with natural ecosystems. The results indicated a significant advancement in the SOS at a rate of -0.077 d/a, a slight delay in the EOS by +0.06 d/a, and a general extension in the LOS by +0.101 d/a across the study area. These shifts led to a persistent increase in VPAIL and a decreasing trend in VPAIV. Spatially, 72.9% of the temperate regions of China exhibited longer senescence periods than green-up phases (VPAIL0), and 97.1% showed higher productivity during the senescence period (VPAIV0). Notably, croplands exhibited significantly higher VPAIV and lower VPAIL than surrounding regions due to a secondary NDVI peak during senescence. Climate driver analysis identified precipitation asymmetry (CAIPrcp) as the dominant control on growing season length asymmetry, with a correlation of R=0.91, explaining 41.4% of its variance. Temperature asymmetry (CAITemp) showed a secondary influence on length asymmetry (R=0.47) but lacked a significant impact on NDVI-based asymmetry (R=-0.39), which was negatively regulated by precipitation. This study provides a comprehensive quantitative characterization of vegetation asymmetry in duration and productivity dimensions over a large spatial scale. With precipitation asymmetry identified as the primary regulator of phenological imbalances in the temperate regions of China, the findings extend the conceptual framework of phenological research beyond simple event-based shifts. The strong spatial consistency revealed by SOM classification underscores the effectiveness of phenological metrics for vegetation type mapping. These insights offer valuable guidance for ecological modeling, carbon cycle estimation, and the development of climate-adaptive ecosystem strategies, including crop management and water regulation. This research emphasizes that understanding intraseasonal dynamics is essential for accurately predicting ecosystem responses to an increasingly asymmetric climate and provides a scientific basis for seasonal resource allocation in environmental management.关键词：vegetation phenology;asymmetry index;climate change;vegetation productivity;remote sensing;NDVI24|10|0更新时间：2026-07-16 -
摘要：Full-polarimetric Synthetic Aperture Radar (SAR) data are highly effective for characterizing the vertical structure of forest canopies and play a crucial role in estimating forest above-ground biomass (AGB). An appropriate polarimetric decomposition method is essential for effectively extracting key polarization features and constructing high-accuracy quantitative inversion models. However, under complex forest canopy conditions, traditional model-free decomposition based on the 3D Barakat degree of polarization is prone to errors caused by depolarization, noise, and variations in observation geometry. These limitations reduce the stability and generalization ability of polarimetric features in AGB retrieval.To address these issues, this study proposed an improved model-free polarimetric decomposition method by replacing the conventional 3D Barakat degree of polarization with the effective degree of polarization (EDoP), which was used to regulate scattering power allocation among different scattering components. L-band full-polarization SAOCOM images from three representative forest regions in Hunan Province, China, were used as data sources. Three decomposition strategies—model-based, traditional model-free, and the proposed EDoP-integrated model-free decomposition—were applied to extract polarization features. Subsequently, forward feature selection was combined with four machine learning regression models, including multiple linear regression, K-nearest neighbor, support vector regression, and random forest, to perform quantitative AGB inversion.Results demonstrate that the EDoP-integrated decomposition significantly improves the power allocation among scattering mechanisms. The proportion of volume scattering decreased from approximately 40% to 30%, while the double-bounce scattering component increased by about 20%, effectively mitigating energy estimation bias caused by noise and depolarization. This adjustment enhanced the physical consistency between scattering components and forest AGB, particularly in heterogeneous canopy structures. Compared with the traditional model-free approach, the proposed method achieved stronger correlations between the extracted polarimetric features and AGB. In quantitative inversion, it increased the coefficient of determination by 0.15—0.30 and reduced the relative root mean square error by 4%—7%. The spatial distribution maps of predicted AGB were also more consistent with actual forest patterns, with less underestimation and overestimation and a weaker saturation effect in high-biomass regions.This study verifies the robustness and transferability of EDoP-integrated model-free decomposition in complex forest environments. The proposed approach effectively suppresses decomposition bias and enhances the accuracy and stability of AGB retrieval. By mitigating the AGB saturation effect and improving model generalization in heterogeneous forest stands, it offers a promising approach for large-scale, high-precision monitoring of forest biomass and carbon storage using polarimetric SAR data.关键词：forest Above-Ground Biomass (AGB);Effective Degree of Polarization (EDoP);model-free decomposition;machine learning;SAOCOM24|25|0更新时间：2026-07-16
Forestry and Agriculture
-
Drone-based remote sensing identification of Himalayan marmot holes integrating habitat information AI导读
摘要：The application of drone imagery combined with deep learning technology has become a crucial approach for identifying Himalayan marmots and monitoring their activity patterns at large spatial scales. Within this workflow, the accurate identification of marmot holes is essential because holes provide a stable spatial proxy for the presence and density of animals. However, current methods are still constrained in practice. Owing to uneven terrain, stones around or above burrow entrances, and soil mounds and other ground objects that partially or completely obstruct openings, a considerable proportion of marmot holes are missed during detection. Consequently, the accuracy and robustness of existing burrow identification techniques remain insufficient for reliable use in plague-related surveillance. This study focused on several typical natural plague foci in the Xizang region. We carried out detailed ground surveys of marmot holes, recording entrance locations, dimensions, and associated habitat characteristics. In parallel, Unmanned Aerial Vehicle (UAV) remote sensing campaigns were conducted to obtain high-resolution aerial imagery over the same sites. On the basis of these observations, we constructed a UAV image dataset that incorporates marmot burrow entrances and surrounding habitat elements such as stone patches and soil mounds formed by excavation. We then designed a data annotation strategy that explicitly integrates habitat information into the labeled objects. Five representative object detection models—YOLOv8, YOLO11, YOLO13, RT-DETR, and RF-DETR—were trained and evaluated under two settings: a traditional annotation method using only the visible burrow entrance and the proposed habitat-integrated annotation strategy. By systematically comparing results from these two strategies, we assessed the effectiveness of incorporating habitat information into the annotation process. Under the traditional annotation method, the accuracy rates for detecting marmot holes achieved by the five models were 93.0% (YOLOv8), 91.6% (YOLO11), 85.5% (YOLO13), 76.5% (RT-DETR), and 83.0% (RF-DETR). When the habitat-integrated annotation strategy was adopted, the performance of most models improved. All models showed higher detection accuracy except for YOLOv8, whose accuracy decreased slightly by 0.2 percentage points. The accuracy gains were 2.2 percentage points for YOLO11, 9.0 percentage points for YOLO13, 16.7 percentage points for RT-DETR, and 6.5 percentage points for RF-DETR. The combination of RT-DETR with the habitat-integrated strategy produced the largest improvement, with an increase of about 17%, while the system integrating YOLO11 with the habitat-integrated strategy achieved the best overall performance. These results show that an annotation strategy integrating habitat information can effectively reduce the risk of missing marmot holes caused by stone cover, soil mounds, and other forms of occlusion that limit traditional annotation methods. By providing rich contextual cues around the entrances, this strategy improves the accuracy and reliability of drone-based identification of Himalayan marmot holes. The proposed annotation framework enhances UAV-based burrow detection, enriches the application of UAVs for investigating plague focuses associated with Himalayan marmots, and offers important implications for monitoring and controlling marmot-borne plague across the Qinghai-Tibet Plateau.关键词：unmanned aerial vehicles;deep learning;data annotation strategy;habitat information;Himalayan marmot holes;plague surveillance156|79|0更新时间：2026-07-16 -
摘要：Conventional pixel offset tracking small baseline subset (SBAS) methods, including Multidimensional SBAS based on pixel offset tracking (PO-MSBAS) and Three-dimensional pixel offset tracking SBAS (3D PO-SBAS), suffer from severe outlier contamination when deriving 3D time-series velocities of surge-type glaciers. These outliers obscure true glacier motion and reduce the reliability of retrieved velocity fields, especially during rapid surge episodes. This study aims to develop a robust signal-processing strategy to identify and suppress outliers effectively, thereby improving the accuracy, continuity, and physical plausibility of 3D time-series velocity monitoring for surge-type glaciers.This study proposes a Robust Trend-Outlier Decomposition (RTOD) algorithm that decomposes an original time series into trend, outlier, and residual components. RTOD first estimates a trend using an adaptive moving average whose window width depends on the series length. RTOD then performs an iterative two-step outlier detection on the detrended residuals: For each time point, it calculates the local median and the Median Absolute Deviation (MAD) within a local window (excluding the point itself and previously identified outliers). It computes two anomaly scores: one based on the raw MAD and the other based on the MAD scaled by 1.4826 (which provides a consistent estimate of the standard deviation). It flags a point as an outlier if either score exceeds a predefined threshold of 3.5. After detecting outliers, the algorithm reconstructs a cleaned signal using a variance-based decision rule: If the residual component has the largest variance contribution, the cleaned signal retains only the trend; otherwise, it retains both the trend and residuals. The RTOD algorithm is integrated into two frameworks: (1) RTOD-PO-MSBAS, where RTOD is applied to the Line of Sight (LOS) and azimuth displacement time series from ascending/descending Sentinel-1A/B synthetic aperture radar (SAR) images before constructing the sparse inversion matrix; and (2) RTOD-3D PO-SBAS, where RTOD is applied to the LOS and azimuth velocity time series after SBAS processing. Both methods then solve for cleaned 3D (east-west, north-south, and vertical) velocity time series. Validation uses 2018–2024 Sentinel-1A/B SAR data (286 descending and 176 ascending images) over the surge-type Osbornebreen Glacier, Svalbard, with cross-validation against Sentinel-2A/B optical velocities and the inter-mission time series of land ice velocity and elevation (ITS_LIVE) annual velocity products.RTOD processing substantially suppresses outliers and improves agreement with independent datasets. On the western branch, compared with Sentinel-2 derived horizontal velocities, the correlation coefficient increases from 0.03—0.77 (traditional methods) to 0.15—0.98 (RTOD-based methods), and the interquartile range (IQR) drops from 35.45—717.26 to 33.59—142.42. On the eastern branch, the correlation coefficient rises from 0.01—0.27 to 0.47—0.80, and the IQR decreases from 140.07—3056.97 to 40.70—113.62. The annual velocity curves from RTOD-PO-MSBAS and RTOD-3D PO-SBAS show strong consistency with ITS_LIVE products, with differences mostly within 150 m/a (western branch) and 100 m/a (eastern branch). Furthermore, the cleaned 3D velocity maps (east-west, north-south, and vertical) exhibit spatially continuous, physically plausible patterns, effectively removing the fragmented, outlier-dominated fields produced by traditional methods.The RTOD method effectively removes outliers from time-series velocity data of surge-type glaciers, substantially enhancing the reliability of 3D time-series monitoring. The proposed RTOD-PO-MSBAS and RTOD-3D PO-SBAS algorithms provide clean, continuous velocity signals while preserving true glacier motion trends. This approach not only supports high-quality studies of surge-type glacier dynamics but also offers a methodological reference for monitoring other large-magnitude surface displacements, such as landslides and debris flows.关键词：surge-type glacier;glacier 3D time-series velocity;robust trend-outlier decomposition;PO-MSBAS;3D PO-SBAS;Osbornebreen Glacier;Svalbard25|13|0更新时间：2026-07-16 -
摘要：Soil moisture is a pivotal variable within the terrestrial hydrological cycle, exerting a critical influence on climate change, agricultural productivity, and ecosystem management. Accurate, high-resolution soil moisture data is essential for applications such as drought monitoring, flood prediction, agricultural precision irrigation, and the refinement of climate and carbon cycle models. However, prevailing soil moisture downscaling methodologies are constrained by significant limitations. This limitations include reliance on single-source soil moisture data products, which fail to leverage the complementary information from observations and models of different scales and physical mechanisms. Furthermore, the selection of predictive covariates is often incomplete, frequently overlooking key hydrological drivers such as terrain-controlled convergence effects, which govern the spatial distribution of moisture. This study addresses these gaps by proposing an integrated downscaling framework. The primary objective is to generate high-resolution, spatiotemporally continuous daily soil moisture data for Northeast China, thereby enhancing the precision and applicability of fine-scale soil moisture monitoring for ecological, hydrological, and agricultural management.This research was conducted in Northeast China, a region characterized by a semihumid continental climate, rolling topography, and significant importance as a national grain production base. The core methodology involved a synergistic downscaling model that incorporated multisource soil moisture products and key terrain convergence features. The model inputs were as follows: (1) Three soil moisture products based on different principles and scales——GLDAS (model assimilation, 25 km), SMAP (satellite passive microwave, 9 km), and ERA5_Land (reanalysis assimilation, ~8 km); (2) A comprehensive suite of covariates encompassing meteorological data (precipitation, air temperature), direct land surface parameters (albedo, evapotranspiration), vegetation indices (NDVI, EVI, NDWI), topographic attributes (elevation, slope, aspect), and, specifically, terrain convergence indices [stream power index (SPI) and topographic wetness index (TWI)] calculated from SRTM DEM; and (3) Soil texture data (sand, silt, clay). All data were harmonized to a 1 km spatial resolution and a daily temporal scale for the period 2019—2023. Four representative machine learning algorithms——random forest (RF), light gradient-boosting machine (LightGBM), deep feedforward neural network (DFNN), and convolutional neural network (CNN)——were employed and comparatively evaluated to construct the downscaling models. Eleven controlled experiments were designed to assess the individual and combined contributions of the multisource soil moisture products and the terrain convergence factors. Model training and validation utilized in-situ soil moisture measurements from 229 stations, with data partitioned into training (190 stations) and validation (39 stations) sets. Model performance was rigorously evaluated using the coefficient of determination (R²), root-mean-square error (RMSE), and Bias.Experimental results demonstrated that the proposed framework successfully enhanced downscaling accuracy. The integration of the three complementary soil moisture products (ERA5-Land, SMAP, and GLDAS) yielded superior performance compared with any single product or dual-product combination, with the best model achieving an R² of 0.849, RMSE of 0.095, and Bias of 0.003. Among the machine learning algorithms, LightGBM consistently delivered the highest accuracy. The introduction of terrain convergence factors, particularly the combined use of SPI and TWI, significantly improved the model’s ability to fit the spatial distribution of soil moisture, raising the R² from 0.843 (without convergence factors) to 0.849. Feature importance analysis confirmed the dominant role of the input soil moisture products, followed by the terrain and hydrological covariates (SPI, TWI, slope aspect). The generated 1 km daily soil moisture product effectively captured fine-scale spatial heterogeneity and seasonal dynamics, showing a high median correlation (r = 0.925) with in-situ validation data, which outperformed the original coarse-resolution products (ERA5_Land: r=0.786, SMAP: r=0.691, GLDAS: r=0.622). Spatial details in low-, medium-, and high-moisture zones were refined, and temporal series trends aligned closely with station observations. A comparative analysis with other recent downscaling studies confirmed the competitive performance of the proposed method.This study developed an effective soil moisture downscaling approach for Northeast China that successfully addresses the limitations of single-data-source dependency and incomplete covariate selection. By synergistically integrating multiscale soil moisture products (ERA5_Land, SMAP, GLDAS) and explicitly incorporating terrain convergence effects (via SPI and TWI), the method generated a high-resolution (1 km) daily soil moisture dataset with improved accuracy and spatiotemporal consistency. The LightGBM algorithm was identified as the most effective model for this task. The results confirm that multisource data integration provides complementary information, and terrain convergence features are crucial for capturing the spatial patterns of soil moisture governed by hydrological processes. The final downscaled product reliably reflects observed soil moisture variations, offering valuable support for fine-scale applications in agricultural water management, hydrological modeling, and ecological monitoring. Future work could focus on enhancing the model’s adaptability to diverse global climates and topographies and incorporating broader temporal data to capture extreme hydrological events.关键词：soil moisture downscaling;Northeast China;multi-source data;machine learning;deep learning;terrain convergence;spatiotemporal continuity;soil moisture products8|11|0更新时间：2026-07-16 -
摘要：Multispectral images are widely applied in land-use classification, but weather conditions have a considerable impact on their quality. Synthetic Aperture Radar (SAR), with its all-weather, all-time capabilities and strong penetration, holds significant application value in land-use classification. However, traditional SAR-based classification methods mainly rely on backscattering features, which restricts improvements in classification accuracy and stability. Compared with land-use classification methods that use single remote sensing images, the fusion of multispectral and SAR images can achieve data complementarity and improve classification accuracy. Hybrid Spectral CNN (HybridSN) is a hyperspectral image classification model based on a hybrid architecture of 3-D and 2-D convolutional neural networks. It effectively fuses spectral and spatial features to improve classification accuracy. However, the HybridSN model tends to misclassify noise as valid features when processing heterogeneous data such as fused SAR and multispectral images, struggling to coordinate the structural features of SAR images with high-level semantic information in multispectral images. To address these issues, we proposed the Hybrid and Recurrent Layer Attention HybridSN (HAR-HybridSN), which introduces a Hybrid Attention Module (HAM) and a Recurrent Layer Attention (RLA) based on HybridSN. The HAM in the proposed model integrates channel and spatial attention. The channel attention module generates channel attention maps through global average and max pooling to highlight discriminative features while suppressing redundant information. After channel aggregation, the spatial attention module then produces spatial attention maps that guide the model to focus on key regions within the image. These two modules work synergistically to enhance the model’s robustness and classification accuracy in complex scenarios such as low signal-to-noise ratio and blurred boundaries. RLA introduces semantic interaction channels between different convolutional layers, efficiently fusing shallow and deep features. Through bidirectional information transmission, low-level details and high-level semantics are co-modeled, which enhances the comprehensive recognition capability of the model for textures, boundaries, and semantic categories. Using GF-3 SAR images and GF-1 multispectral images as data sources, we conducted qualitative and quantitative evaluations of the classification results of the HAR-HybridSN model in comparison with six models, including HybridSN. Results show that the overall accuracy, average accuracy, and kappa coefficient of HAR-HybridSN are higher than 93%, 91%, and 91%, respectively. In areas with mixed ground objects and blurred boundaries, it can accurately identify key semantic features and effectively suppress interference from redundant information, leading to enhanced discriminative ability for complex ground objects. Specifically, in complex ground objects, the overall accuracy of the HAR-HybridSN model is about 1% higher than that of other models, and it maintains better consistency and continuity in categories such as buildings and cultivated land. In terms of running time, HAR-HybridSN maintains high classification accuracy while exhibiting favorable computational efficiency, with its overall performance surpassing that of other models. Furthermore, under small-sample conditions, it demonstrates robust classification capability and generalization performance.关键词：multispectral images;polarimetric SAR;Multi-scale Features;HAR-HybridSN;land use classification10|13|0更新时间：2026-07-16 -
摘要：Fine-grained ship detection in optical remote sensing imagery is an important research topic in maritime monitoring and ocean observation. Compared with coarse-grained ship detection, fine-grained ship detection aims to distinguish specific ship types with highly similar visual appearances, which places higher demands on dataset scale, category diversity, and annotation accuracy. However, existing ship detection datasets are generally limited in the number of fine-grained categories, instance scale, and scene diversity, which restricts the performance and generalization ability of deep learning-based oriented object detection models. The objective of this study is to construct a large-scale fine-grained ship detection dataset and expand data diversity through controllable synthetic data generation, thereby providing a reliable data foundation and benchmark for fine-grained ship detection in optical remote sensing imagery.First, a large-scale fine-grained ship detection dataset, named LAFI, is constructed using high-resolution optical remote sensing images collected from 36 representative ports worldwide. The dataset contains 8,000 images acquired under diverse imaging conditions and complex maritime environments. A total of 49 fine-grained ship categories are defined, and 48,717 ship instances are manually annotated using oriented bounding boxes to accurately describe ship orientation and geometric characteristics. Second, to alleviate the limitations of real data in terms of scale and scene coverage, this study designs a controllable diffusion-based data generation framework to extend LAFI into a million-scale synthetic dataset, referred to as LAFI-Diffusion. The diffusion model is guided by incorporating structured textual prompts that describe scene types, weather conditions, and temporal information to generate realistic ship images under diverse maritime scenarios. The generated synthetic samples are further filtered and combined with real data to form large-scale training sets suitable for fine-grained ship detection. Finally, several representative oriented object detection methods are selected and evaluated on the constructed datasets to analyze the effectiveness of synthetic data augmentation and establish benchmark results.Experimental results show that the proposed LAFI dataset provides improved category richness, instance scale, and scene diversity compared with existing fine-grained ship detection datasets. Incorporating synthetic data from LAFI-Diffusion into the training process consistently enhances detection performance and generalization ability across different oriented object detection models. Performance gains are particularly evident in complex maritime environments, such as crowded ports and scenes with varying sea states and illumination conditions. Benchmark evaluations indicate that the contribution of synthetic data varies across detection methods, suggesting that appropriate integration strategies are important for fully exploiting synthetic data.This paper presents a large-scale fine-grained ship detection dataset that integrates real optical remote sensing imagery with controllable diffusion-based synthetic data generation. By substantially expanding data scale and enhancing scene diversity, the proposed LAFI-Diffusion dataset effectively addresses the limitations of existing fine-grained ship detection benchmarks. The experimental results confirm that synthetic data can serve as an effective complement to real-world samples, improving detection accuracy and robustness for oriented ship detection models. The released datasets and benchmark results provide valuable support for future research on fine-grained ship detection and related remote sensing applications.关键词：remote sensing imagery;ship object detection;dataset;diffusion model;synthetic data generation;Fine-grained recognition;oriented bounding box;data augmentation380|644|0更新时间：2026-07-16 -
摘要：This study aimed to systematically evaluate the accuracy and consistency of the Aerosol Optical Depth (AOD) product from the medium resolution spectral imager-II (MERSI-II) onboard the FY-3D satellite. Cross-comparisons with global ground-based observations and international satellite products provided a scientific basis for algorithm optimization and product application.The MERSI-II AOD product from 2019 to 2023 was compared with observations from 538 global AERONET sites and three international satellite AOD products: Aqua/MODIS, NOAA-20/VIIRS, and Suomi-NPP/VIIRS. Spatiotemporal matching was applied, and statistical metrics, including the correlation coefficient (R), Mean Bias (MB), and Root Mean Square Error (RMSE), were used to analyze product accuracy and bias characteristics.MERSI-II AOD showed good agreement with AERONET observations (R=0.822, MB=0.048, and RMSE=0.169), with 60.32% of samples falling within the expected error margin. Compared with other satellite products, all four exhibited similar temporal trends: MERSI-II AOD performed best overall in 2023 and achieved high accuracy in winter (November-February, R0.83, MB0.045). Spatially, all products showed highly similar distribution patterns, with high correlation over densely vegetated regions (e.g., eastern North America, R≥0.88) but poor performance over bright surfaces such as North Africa and the Middle East. MERSI-II AOD exhibited an average positive bias of 0.048 relative to AERONET, similar to Suomi-NPP/VIIRS (0.056). Further bias analysis indicated that the MERSI-II AOD bias was negatively correlated with aerosol loading and scattering angle but positively correlated with shortwave infrared NDVI, solar zenith angle, and viewing zenith angle.The overall accuracy of MERSI-II AOD was comparable to that of Suomi-NPP/VIIRS but slightly lower than those of NOAA-20/VIIRS and MODIS. Performance requires further improvement under high pollution conditions and over bright surfaces (e.g., urban and bare soil areas). Future algorithm enhancements will focus on incorporating regional aerosol models, optimizing the surface reflectance model for bright surfaces, and improving the timeliness of radiometric calibration updates. These coordinated efforts in algorithm and data processing will enhance product accuracy, thereby strengthening the contribution of the Fengyun satellite series to the global aerosol observation system.关键词：Fengyun-3D (FY-3D);medium resolution spectral imager-II (MERSI-II);Aerosol Optical Depth (AOD);ground-based validation;error analysis10|12|0更新时间：2026-07-16 -
A spatiotemporal-aware and cascaded alignment framework for remote sensing image change detection AI导读
摘要：High-resolution remote sensing image Change Detection (CD) identifies land-cover changes from bi-temporal images of the same geographic area and is essential for urban monitoring, disaster assessment, ecological surveillance, and land resource management. However, practical CD remains challenging because real bi-temporal images are often affected simultaneously by temporal appearance discrepancies and spatial geometric misalignment. Illumination changes, seasonal variations, shadows, and radiometric inconsistencies may produce spectral pseudo-changes, while viewpoint differences and imperfect registration may introduce building displacement, boundary distortion, and local spatial offsets. Although these two types of interference are physically different, they become coupled during feature extraction, differencing, alignment, and fusion, causing error propagation and making it difficult to separate true semantic changes from irrelevant variations. To address this problem, this study proposes a Spatiotemporal-Aware Multi-scale Feature Flow (SAMFF) network for robust remote sensing image CD under compound spatiotemporal interference.The proposed SAMFF follows a cascaded alignment strategy that decomposes compound interference suppression into four stages: temporally consistent feature representation, spatial geometric correction, multi-scale enhancement, and confidence-guided fusion. First, a cross-temporal attentional alignment module (CTAAM) is embedded in the feature extraction stage to enable deep interaction between bi-temporal features. By combining cross-temporal attention with local context-aware and channel semantic enhancement branches, CTAAM aligns channel semantic and local contextual representations, thereby reducing pseudo-change responses caused by illumination and seasonal variations. Second, a flow-aligned difference module (FADM) is introduced to handle geometric mismatch. It first enhances temporally aligned features through multi-scale channel attention and then estimates a dense differential flow field to warp and align bi-temporal features at multiple scales. This design improves pixel-level correspondence and suppresses false responses caused by registration errors. Third, a Cross-Scale Attention Feature Enhancement (CSAFE) module aggregates multi-scale contextual information through a dilated convolution pyramid and top-down attention-based feature interaction, strengthening the discriminative representation of real change regions. Finally, a Confidence Flow Guided Feature Fusion Module (CFGFFM) acts as the decoder. It estimates confidence-aware flow fields and adaptively fuses high-level semantic features with low-level spatial details, reducing residual noise in low-confidence regions and producing a spatially consistent change map.Experiments are conducted on three public remote sensing CD datasets, namely, SYSU-CD, WHU-CD, and GZ-CD. SAMFF achieves F1/IoU of 83.83%/72.16%, 93.44%/87.69%, and 88.40%/79.23%, respectively, on the three datasets, outperforming representative CNN-, Transformer-, and alignment-based comparison methods in overall CD accuracy. Qualitative results further show that SAMFF can better suppress pseudo-changes caused by roof color differences, shadows, vegetation appearance variations, building parallax, and local misregistration. Ablation studies verify the effectiveness of CTAAM, FADM, CSAFE, and CFGFFM, demonstrating that the proposed temporal-to-spatial cascaded alignment design consistently improves performance. In addition, SAMFF achieves a good balance between accuracy and efficiency, with 14.28 M parameters and 3.65 G FLOPs for 256×256 image inputs.SAMFF provides an effective cascaded framework for high-resolution remote sensing image CD by explicitly addressing the coupled influence of temporal appearance discrepancy and spatial geometric misalignment. Instead of directly learning changes from mixed noisy features, the proposed method progressively suppresses temporal pseudo-changes, corrects geometric offsets, enhances multi-scale change representations, and fuses features with confidence guidance. Experimental results confirm its robustness, accuracy, and deployment potential in complex bi-temporal remote sensing scenarios.关键词：change detection;remote sensing;spatiotemporal awareness;feature alignment;multiscale fusion;deep learning15|12|0更新时间：2026-07-16
Models and Methods
- Postal code：100190
- Tel：010-58887052 Email：nrsb@aircas.ac.cn
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备20021838号-9
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0