Issue Ebookcover storyIssue CoverIssue Contents
Latest Issue
Online FirstList of IssuesDocument Types
2026
Issue 7
Issue 6
Issue 5
Issue 4
Issue 3
Issue 2
Issue 1
2025
2024
2023
2022
2021
2020
2019
2018
2017
Abstract：The scientific delineation of near-space boundaries is a fundamental prerequisite for the orderly development of related industries. Persistent inconsistencies in historical demarcation criteria and a mismatch between traditional definitions and current technological needs have left the concept of near space ambiguous. This study aims to transcend the limitations of the traditional Kármán line by proposing a new, multidimensional partition framework. The objective is to provide a coherent, scientifically grounded definition that integrates atmospheric physics, engineering constraints, and airspace management principles to guide future research, technology development, and rulemaking.This study proposes a three-segment partition framework for near space—the aerodynamic domain (lower near space: 18—80 km), the transition domain (middle near space: 80—200 km), and the perturbation domain (upper near space: 200—300 km)—based on the continuous variation in the atmospheric environment and its consequent design constraints on flight platforms. The methodology integrates prior knowledge from three distinct perspectives: (1) natural science, analyzing the continuous evolution of atmospheric physics and space environment from the stratosphere to the ionospheric F-layer peak; (2) engineering and technology, evaluating typical flight platforms (solar-powered drones, stratospheric aerostats, hypersonic vehicles, suborbital vehicles, and very low Earth orbit satellites) to determine operational boundaries; and (3) airspace management, assessing the framework’s value in bridging current air traffic control systems and resolving legal ambiguities between air and space.The study’s findings corroborate the framework’s scientific validity across multiple dimensions. (1) The 18—300 km range represents a complete Earth system characterized by gradual atmospheric parameter changes and a continuous evolution of dominant physical processes—from continuum flow aerodynamics, through an aerodynamics–orbit coupling transition, to a rarefied atmosphere perturbation regime. (2) Engineering analysis of five platform types indicates that 18 km is the feasible lower boundary for long-endurance flight (the upper limit of conventional aviation), and 300 km is the practical upper boundary where atmospheric drag becomes the decisive perturbation factor on orbital lifetime. This framework seamlessly connects the continuum from aviation to astronautics. (3) The three-tier partition provides a clear technical benchmark for bridging the current aviation management framework and addressing legal ambiguities in outer space.This definition of near space serves as a research subject that delineates the operational scope for future industry development. For technology development, the framework, grounded in physical mechanisms, helps the industry avoid blind exploration by forming a stepwise R&D roadmap from “aviation aerodynamics-dominated” to “astronautic orbit-dominated” technologies. For standard setting, it facilitates differentiated airworthiness certification, environmental adaptability, and communication standards for each of the three layers. For international cooperation, the “lower–middle–upper near-space” framework enables a new paradigm of “layered collaboration,” allowing nations to select zones matching their industrial capabilities. Therefore, this study does not merely extend the current 18—100 km cognitive range; by establishing the “aerodynamic–transition–perturbation domain” framework, it provides a theoretical foundation for the systematic understanding and differentiated utilization of near space from an environment–mission coupling perspective.Keywords：Near-space;space demarcation;natural science;engineering technology;airspace management550|189|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Scholar's View Point
Review of deep learning-based spatiotemporal fusion of multisource remote sensing images AI Introduction
Abstract：With the continuous advancement of sensor technology and earth observation capabilities, remote sensing data have become a fundamental information resource supporting scientific research and socioeconomic development, and the demand for high-spatiotemporal-resolution remote sensing imagery has become increasingly urgent. However, constrained by the inherent trade-off between scanning swath width and pixel size, a single satellite sensor can hardly obtain remote imagery with high spatial and temporal resolutions simultaneously. This limitation creates a dilemma in earth observation tasks—such as vegetation phenology monitoring, crop growth monitoring, disaster assessment, urban planning, water quality monitoring, and land cover change detection—where data often offer “high timeliness but lack of detail” or “high precision but lack of continuity.” Consequently, this constraint significantly limits the application potential of remote sensing imagery.Spatiotemporal fusion of multisource remote sensing imagery offers a viable solution by integrating data from different sensors. By blending high-spatial/low-temporal resolution images with low-spatial/high-temporal ones, this approach mitigates the uncertainty and incompleteness inherent in a single data source. The resulting fused imagery achieves high spatial and temporal resolutions simultaneously. This approach effectively alleviates the trade-off between spatial and temporal resolutions in a single satellite sensor. It also meets the practical demand for remote sensing data with high spatial and temporal resolutions.However, owing to the gap between physical models and real-world conditions, unreasonable prior assumptions, and reliance on handcrafted features, traditional fusion methods still suffer from limited performance when handling highly nonlinear relationships or large-scale multisource data. Some methods also perform poorly in heterogeneous regions and struggle to effectively fuse multisource remote sensing data across spatiotemporal scales. The emergence of deep learning has introduced advanced spatiotemporal fusion techniques, such as convolutional neural networks and diffusion models. These approaches offer strong representation learning capability, high adaptability, and end-to-end modeling advantages. They can effectively extract multisource spatiotemporal features and provide a new paradigm for spatiotemporal fusion.To keep pace with recent advances in this field, this study provides an in-depth analysis of deep learning-based multisource remote sensing image spatiotemporal fusion methods in accordance with their underlying principles and characteristics. These methods are categorized into five major classifications, namely, convolutional neural networks, transformers, generative adversarial networks, recurrent neural networks, and diffusion models. Subsequently, commonly used open-source datasets, deep learning algorithms, and performance evaluation metrics for remote sensing image spatiotemporal fusion are summarized and introduced. Curated links to datasets and model codes are also provided to support future research in this field.In this study, state-of-the-art methods, including traditional and deep learning-based approaches, are selected for quantitative and qualitative evaluations across the CIA, DX, and SW datasets. The spatiotemporal fusion algorithms are validated, and a comparative analysis of their respective strengths and limitations is conducted on the basis of the fusion results. Finally, the development of various types of spatiotemporal fusion algorithms and the results of experimental analyses are synthesized to summarize the main challenges faced in the field of spatiotemporal fusion of remotely sensed images at the present stage and the potential future development directions.Keywords：spatiotemporal fusion;multi-source remote sensing images;deep learning;algorithmic principles;technique classification;datasets521|151|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Challenges， methods， and opportunities in feature extraction of visible light and SAR images AI Introduction
Abstract：Visible light and Synthetic Aperture Radar (SAR) images are two primary data sources in satellite remote sensing, widely used in fields such as urban planning, disaster prevention, and national security. Feature extraction serves as a key bridge connecting remote sensing images with high-level applications, directly influencing the effectiveness of intelligent interpretation in complex scenarios. Feature extraction methods have undergone a profound transformation from handcrafted design to data-driven learning, achieving significant breakthroughs, particularly with the advancement of deep learning. However, due to the high cost and strong domain specificity of remote sensing data annotation, previous data-driven methods still face challenges such as insufficient generalization and limited interpretability. To address these issues, existing researches have proposed a “data + knowledge” dual-driven framework, which leverages data-driven perception and knowledge-driven reasoning to achieve higher-level intelligence, aiming to break through the bottlenecks of generalization and interpretability inherent in single data-driven paradigms. To reveal the technical logic and core value of this dual-driven framework in feature extraction for visible light and SAR images, this paper first systematically reviews traditional feature extraction methods and deep learning-based methods, summarizing the main challenges faced by these two categories. It then elaborates on knowledge-driven feature extraction, categorizing the knowledge used to guide networks into visual knowledge, geospatial knowledge, and physical knowledge, and discusses the training paradigms for knowledge-guided networks. Finally, it outlines future trends, identifying the construction of remote sensing foundation models, the dual-driven development of data and knowledge, and the cross-integration of cutting-edge technologies as key research directions.This study follows a structured review methodology, focusing on literature from authoritative sources such as IEEE TGRS, ISPRS Journal, CVPR, and ICCV between 2019 and 2025. First, traditional feature extraction methods—including edge detection (e.g., Sobel, Canny, ROEWA), texture analysis (e.g., Gabor filters, GLCM), color features, and keypoint extraction (e.g., SIFT, Harris)—are summarized and analyzed for their strengths and limitations. Second, deep learning-based methods are examined, covering convolutional neural networks for local spatial modeling, graph neural networks for relational reasoning, and vision transformers for global context understanding. Third, a knowledge-driven framework is proposed to address challenges such as semantic ambiguity, scale variation, imaging condition sensitivity, and speckle noise interference. This framework integrates three categories of knowledge—visual knowledge (e.g., edges, textures), geospatial knowledge (e.g., DEM, topological relations), and physical knowledge (e.g., scattering models, atmospheric physics)—into deep networks through mechanisms such as knowledge as architecture, knowledge as data, and knowledge as loss.The analysis demonstrates that traditional methods offer high interpretability and computational efficiency but lack adaptability to complex scenes and nonlinear relationships. Deep learning methods significantly improve feature representation and task performance but are heavily dependent on large annotated datasets and suffer from poor generalization under domain shifts. The proposed knowledge-driven approach effectively mitigates these issues. For example, incorporating edge prior knowledge enhances boundary precision in semantic segmentation, embedding elevation data improves terrain-aware classification, and integrating physical models such as electromagnetic scattering models aids in SAR image interpretation and speckle suppression. Experimental results from cited studies show performance gains across various tasks, such as increased mIoU in segmentation and enhanced accuracy in object detection, confirming that knowledge-guided features are discriminative, stable, and interpretable.Feature extraction for visible light and SAR images has evolved from manual design to data-driven learning and is now progressing toward a hybrid paradigm that combines data-driven perception with knowledge-driven reasoning. This study highlights that integrating visual, geospatial, and physical knowledge into deep learning models can address key challenges in remote sensing feature extraction, including semantic ambiguity, scale extremes, imaging variability, and noise interference. Future research should focus on developing large-scale remote sensing foundation models, deepening the synergy between data and knowledge, and promoting cross-disciplinary integration with emerging technologies like vision-language and diffusion models. Such advancements will further enhance the intelligence, reliability, and applicability of remote sensing image interpretation systems.Keywords：remote sensing intelligent interpretation;visible light images;SAR Images;manual feature extraction;depth feature extraction;data and knowledge driven491|2561|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Reviews
Abstract：Against the backdrop of China’s “dual carbon” strategic goal, accurately characterizing the spatiotemporal dynamics of industrial heat sources (IHS) is critical for advancing green, high-quality, and sustainable industrial development. As the core proxy indicator of energy-intensive industrial activities, IHS directly reflect the intensity of production operations and associated carbon emissions, making them a key monitoring target for low-carbon industrial transition. However, existing research lacks dynamic IHS datasets with large spatial coverage, long time series, and high spatiotemporal resolution. Most related studies either focus on short-term cross-sectional snapshots or lack fine-grained industrial category information, failing to capture interannual industrial evolution and sector-specific emission patterns. This gap has hindered effective tracking of national industrial policy effects, such as structural adjustment and backward production capacity reduction, driving the need for a long-term, refined IHS monitoring framework.To fill this research gap, this study adopts long-time-series 375 m resolution active fire hotspot (ACF) product data derived from the Suomi National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP VIIRS). On this basis, an improved K-means clustering algorithm is proposed for reliable industrial heat source extraction. Combined with geographic Point of Interest (POI) topological correlation analysis and fine feature interpretation of high-resolution remote sensing images for typical factories and mining enterprises, we conduct systematic identification and fine classification of national industrial heat sources during 2021—2023. Following the standardized technical framework and complete data processing workflow, the research process mainly includes four core links: remote sensing data preprocessing, industrial heat source identification based on improved adaptive Kmeans algorithm, IHS classification relying on POI topological correlation analysis, and final result statistics and comprehensive analysis.For the first time, we constructed and openly released a standardized vector-format national IHS dataset covering 2012—2023 with complete industrial category attributes, which is freely accessible to global researchers, filling the long-standing gap of high-resolution, category-labeled IHS data for China that can support multi-year policy assessment. Experimental results show that the improved method maintained a high overall recognition accuracy of 98.14%, while achieving comprehensive optimization in the quantity of identified targets, recognition precision, average spatial particle size, and regional spatial coverage. The established dataset integrates 20 key characteristic parameters, including factory geographic coordinates, annual production operation status, and industrial classification attributes, fully recording the radiation flux features and production activity intensity of diverse IHS types.The findings indicate that extending the research time span to the 12-year period, coupled with POI-topology-based IHS category recognition, can effectively reveal the differentiated spatiotemporal evolution patterns of various IHS during the critical period of China’s industrial transformation and upgrading. The established dataset integrates 20 key characteristic parameters offers refined data support for subsequent accurate industrial carbon emission accounting, industrial spatial layout analysis, and regional socio-economic development evaluation and evidence-based policy formulation.Keywords：industrial heat source;active thermal anomaly data;long time sequence;remote sensing165|43|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Data Article
PTV3-BLSeg： A branch-leaf segmentation model for fine-scale structural analysis of tree point clouds AI Introduction
Abstract：Tree point cloud segmentation plays a crucial role in forest structural analysis, 3D forest reconstruction, and the accurate estimation of tree parameters such as diameter at breast height, crown structure, and biomass. However, automatic branch-leaf separation from tree point clouds remains a challenging task. Major difficulties include the ambiguity of boundaries between trunks, branches, and foliage, the severe class imbalance caused by the significantly larger number of foliage points compared with that of trunk points, and the complex fine-scale geometric structures present in tree crowns. These challenges often lead to misclassification of trunk points as foliage and inaccurate representation of structural details, which ultimately limits the reliability of downstream forestry applications. Therefore, developing a robust and accurate segmentation method capable of effectively handling boundary ambiguity, class imbalance, and fine-grained structural representation is essential for improving the quality of tree point cloud analysis.To address these issues, this study proposes Point Transformer V3 (PTV3)-BLSeg, an improved tree point cloud segmentation model based on the PTV3 architecture. The proposed model introduces three key improvements. First, fusion skip connections are incorporated into the decoder to strengthen cross-level feature propagation and enhance the interaction between shallow geometric features and deep semantic features, thereby improving the model’s ability to capture local structural details and global contextual information. Second, a boundary-enhanced loss function is designed to emphasize points located near trunk-leaf boundaries by assigning them higher training weights. This strategy helps the model focus on subtle structural transitions and improves discrimination between adjacent classes. Third, a deep supervision mechanism is introduced by adding auxiliary supervision to intermediate layers, which guides the feature learning process at multiple depths, stabilizes network training, and improves the learning efficiency of trunk-related features that are typically underrepresented in the dataset. The proposed model was evaluated using three different tree point cloud datasets representing diverse forest environments: the Smart Tree Dataset, the Genhe Dataset from Inner Mongolia, and the Hengzhou Dataset from Guangxi.Experimental results demonstrated that the proposed PTV3-BLSeg model consistently outperformed the baseline PTV3 model across all evaluation datasets. On the Smart Tree Dataset, PTV3-BLSeg achieved an overall accuracy of 93.33%, with trunk and leaf F1-scores of 75.61% and 96.09%, representing improvements of 1.36%, 4.14%, and 0.85%, respectively, compared with the baseline. On the Genhe dataset, the model achieved an overall accuracy of 83.83%, with trunk and leaf F1-scores of 60.77% and 89.50%, improving by 5.04%, 12.86%, and 9.13%, respectively. On the Hengzhou Dataset, the model obtained an overall accuracy of 81.87%, with trunk and leaf F1-scores of 71.32% and 86.60%, which corresponded to improvements of 4.49%, 12.73%, and 1.48%, respectively. These results indicate that the proposed model significantly enhances trunk recognition performance while maintaining high accuracy in leaf segmentation.Overall, PTV3-BLSeg effectively addresses key challenges in tree point cloud segmentation, including boundary ambiguity, class imbalance, and the difficulty of recognizing fine-scale structures. By integrating fusion skip connections, a boundary-enhanced loss function, and a deep supervision strategy, this model improves segmentation accuracy and training stability. The consistent performance improvements observed across multiple datasets demonstrate the robustness and generalization capability of the proposed approach. This study provides an effective and reliable solution for high-precision tree point cloud segmentation and offers valuable technical support for advanced forestry applications such as 3D forest inventory, tree structure analysis, and ecological monitoring.Keywords：branch-leaf separation;Point Transformer V3;boundary enhancement;deep supervision;point cloud semantic segmentation;skip connection;Feature fusion;3D point clouds139|39|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Abstract：Mountain forests, as critical terrestrial carbon reservoirs, exert significant impacts on atmospheric chemical processes, landscape ecological functions, and biodiversity through key climate feedback mechanisms. However, owing to the constraints of complex topography and observation conditions, the accurate characterization of tree species composition and spatial distribution in mountain forests remains challenging. The integration of terrestrial satellite monitoring technologies, represented by Sentinel-2, with artificial intelligence algorithms has greatly transformed traditional forest survey approaches, offering new opportunities for tree species monitoring and fine-scale mapping. In this study, ground sample plots and multisource remote sensing data were integrated, and multitemporal Sentinel-2 imagery from 2020 to 2023 was processed via the Google Earth Engine platform. A multidimensional feature optimization scheme was developed using the Jeffries-Matusita (JM) distance, followed by SNIC segmentation and machine learning approaches to build a random forest model with optimal features for fine-scale tree species classification in the Qinling-Daba Mountains, thereby revealing spatiotemporal differentiation patterns. The results are as follows: (1) Topographic features are key variables for tree species identification in the Qinling-Daba Mountains, and JM distance-based feature combination schemes can effectively improve feature selection. (2) Under conditions of limited samples, medium-resolution imagery, and large-scale areas, the object-oriented random forest model based on SNIC segmentation demonstrated robust classification performance, with an average overall accuracy of 0.837, a kappa coefficient of 0.791, and a mean F1-score of 0.799. (3) Quercus, Pinus, and Cunninghamia species were widely distributed, whereas other species exhibited scattered distributions. (4) From 2020 to 2023, the forest area showed a slight upward trend, indicating continuous ecological improvement in the region. Pinus and Quercus contributed the most to forest change, whereas species with smaller area proportions exhibited more frequent transitions. This study provides valuable data support for carbon sink assessment and biodiversity conservation in complex mountain forest ecosystems.Keywords：mountain forests;tree species identification;spatial distribution dynamics of tree species;remote sensing data;spatiotemporal differentiation of tree species189|52|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21A deep learning-based early forest fire detection method integrating visible light and thermal infrared images from unmanned aerial vehicles AI Introduction
Abstract：Forest fires pose serious threats to ecosystems, human life, and property. Detecting fires during the incipient combustion stage is essential for rapid suppression and for preventing escalation into catastrophic events. UAV-based remote sensing provides flexible deployment, broad spatial coverage, and high spatial resolution, making it highly suitable for forest fire monitoring. However, early fire detection remains challenging because (1) incipient flames and smoke are small and easily confused with complex forest backgrounds; (2) dense vegetation can occlude targets; (3) illumination changes and environmental variability produce blurred boundaries and unstable target morphologies; and (4) limited onboard computing resources restrict the use of complex detection models. This study develops a lightweight and efficient detection model that integrates RGB and thermal infrared imagery to enable accurate real-time detection of early-stage flames and smoke on resource-constrained UAV platforms.We propose EFFNet (early forest fire net), a dual-modal deep learning model for simultaneous flame and smoke detection. The architecture contains four components designed for UAV remote sensing and the visual characteristics of early fires. (1) An EF (early fusion) module at the input layer combines RGB texture cues with thermal infrared temperature patterns, using their complementary information to improve robustness under nighttime conditions, smoke interference, and occlusion. (2) A HGRNet (heterogeneous global routing network) is used as the backbone; it applies parallel convolutions with different kernel sizes to enlarge receptive fields and capture multi-scale diffusion patterns of flames and smoke, while local max-pooling preserves fine-grained shallow features required for small-target detection. (3) A FAN (Feature Aggregation Network) constructs dual top-down fusion paths and adds a high-resolution detection branch to enhance feature representation and improve localization accuracy for small targets. (4) An ESCHead (efficient semi-coupled detection head) adopts a semi-decoupled design with shared feature extraction and 1×1 convolution-based dimensionality reduction, reducing computational cost while improving alignment between classification and localization branches and thereby decreasing false alarms.Comprehensive experiments on the RGBT-3M dataset validate the effectiveness of EFFNet. The model achieves mAP50 of 97.2%, mAP50—95 of 68.3%, precision of 94.9%, and recall of 94.1%. It contains only 1.84 M parameters and requires 5.6 GFLOPs of computation. Real-time evaluation yields an inference speed of 362.72 frame pairs per second and a single-pair processing time of 2.757 ms. Visualization results demonstrate that EFFNet accurately detects small-scale flame and smoke targets in complex scenarios, including vegetation occlusion, nighttime imaging, and blurred object boundaries, while maintaining low false-positive and false-negative rates.Overall, EFFNet achieves a favorable balance among detection accuracy, computational efficiency, and real-time performance, satisfying the operational requirements of UAV platforms for early forest fire monitoring. Its lightweight architecture supports practical deployment on resource-limited edge computing devices, and its high precision and reliability provide technical support for early warning and rapid response in forest fire management systems.Keywords：UAV;remote sensing image;deep learning;object detection;early forest fire;object detection process;dataset;performance evaluation metrics145|48|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Forestry and Agriculture
Accuracy assessment and analysis of ICESat-2 Altimetry Data in urban areas based on multiple factors AI Introduction
Abstract：ICESat-2 satellite laser altimetry data have the advantages of high accuracy, high temporal resolution, and low cost, making them a valuable data source for urban 3D modeling. However, systematic accuracy assessments of these data in urban environments remain limited.This study designed and implemented a methodology to improve the accuracy of ICESat-2 altimetry data using airborne LiDAR data and to evaluate the performance of the corrected data. On the basis of the ATL03 and ATL08 products, a complete urban spaceborne altimetry dataset was first generated for the study area. Elevation outliers were removed, and quality screening was conducted on the basis of multiple attributes of the original ICESat-2 data. Subsequently, airborne LiDAR point clouds from one part of the study area were used to perform horizontal and vertical geolocation offset correction on the spaceborne altimetry data. The accuracy of the corrected ICESat-2 data was then evaluated in another part of the study area. Finally, data accuracy was analyzed in consideration of multiple factors, including urban land cover types, strong and weak beams, acquisition time, acquisition season, terrain slope, and fractional vegetation cover.Result Results showed that the original ICESat-2 altimetry data in the study area exhibited a slight underestimation (ME=-0.10 m), with MAE, RMSE, and LE90 of 0.67, 1.57, and 1.46 m, respectively. Specifically, spaceborne laser data over construction land and other land cover categories demonstrated relatively low accuracy, whereas data over transportation areas, building ancillary land, and agricultural and forestry land exhibited higher accuracy. After correction, the systematic bias and long-tail errors of the ICESat-2 altimetry data were effectively reduced, and the overall accuracy was significantly improved (MAE=0.48 m, RMSE=1.11 m). In particular, the accuracy of data acquired by weak beams and during daytime was enhanced significantly. Data acquired in spring and winter, over low-slope terrain and areas with high vegetation coverage, exhibited higher accuracy.This study demonstrates that local high-accuracy airborne LiDAR data can be used to correct ICESat-2 altimetry data and effectively improve their accuracy in adjacent regions. The usability of daytime and weak-beam ICESat-2 data is enhanced, and the number of high-quality photons increases accordingly. This work can support urban 3D modeling and information fusion using ICESAT-2 data.Keywords：ICESat-2;spaceborne LiDAR altimetry;data correction;accuracy assessment;multiple factors;urban three-dimensional modeling115|45|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Abstract：Land use represents one of the most intensive forms of human modification of the Earth's surface. It has both economic and natural attributes, and future populations are expected to become increasingly concentrated in urban areas. In the context of rapid urbanization and diversified functional demands, automatic and efficient acquisition of urban land use information is essential for the design, construction, and management of sustainable urban environments. Traditional data-driven algorithms are limited by “black-box” opacity, strong dependence on training data, and difficulty incorporating prior knowledge and semantic information. Reliance on remote sensing data alone also makes accurate land use identification difficult. In addition, existing studies of urban land use often neglect the explicit representation of category semantics, and inconsistencies in semantic representation across classification systems can create a semantic gap between data users and data producers. To obtain precise urban land use information, this study proposes a semantic-guided ontology knowledge-driven urban land use classification.To address these issues, this study develops a semantic-guided ontological knowledge-driven approach that mitigates the limitations of purely data-driven methods. The method integrates multi-source data and products, including remote sensing imagery, POI, AOI, OSM, LCZ, and WSF data. First, land use types are semantically reconstructed using the EAGLE matrix, and land use semantics are decomposed into ontological primitives. Next, crowdsourced-data features, remote sensing indices, and morphological characteristics are extracted to match the attribute features of the primitives, and an ontological model is constructed from these primitives. A segmentation strategy that integrates spectral information and road network structure is then used to generate homogeneous parcels as instances for ontological reasoning. During ontological model construction, constraint rules are established by defining relationships between primitives and data features, enabling land use types for each parcel to be determined through rule-based reasoning. To validate the method, the highly complex downtown area of Toronto, Canada, is selected as the study area. Land use classification is performed according to the “Current Land Use Classification” (GB/T 21010-2017), with an additional “mixed-use” category introduced for practical application.The experimental results validate the effectiveness of the proposed knowledge-driven framework. The method achieves an overall classification accuracy of 87.02% and a Kappa coefficient of 0.84, representing a substantial improvement over two baseline data-driven approaches: support vector machine (SVM) and the MobileNet deep learning network. A major advantage of the proposed method is its strong interpretability, which derives from the transparent reasoning process. In addition, the EAGLE matrix helps bridge the semantic gap between users’ conceptualization of land use and its representation in data products. The proposed fusion segmentation strategy, which integrates spectral and road network information, delineates urban land use boundaries that are more consistent with physical reality and functional units.The knowledge-driven approach proposed in this study mitigates key limitations of data-driven methods in urban land use classification, including weak interpretability, limited accuracy, and semantic ambiguity. By integrating multi-source data, the EAGLE matrix, and ontology, the framework provides a reliable solution for representing the complex characteristics of urban environments. This research establishes a foundation for future studies that aim to improve the intelligence and sustainability of urban planning and management.Keywords：knowledge-driven;semantic information;EAGLE matrix;multi-source data;urban land use96|45|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Impacts of rural settlement evolution on regional carbon storage in the Yellow River Delta based on multisource data AI Introduction
Abstract：Since the implementation of China’s Reform and Opening-up policy, the continuous advancement of industrialization and urbanization has profoundly reshaped rural spatial patterns. In many rural regions, a typical phenomenon characterized by “population decrease but land increase” has emerged, resulting in the disorderly expansion and hollowing of rural settlements. This process has posed significant challenges to rational land resource utilization, ecological environment protection, and food security. Existing studies have mainly focused on the spatial pattern evolution and driving mechanisms of rural settlements, while relatively limited attention has been paid to the carbon storage changes induced by the orderly withdrawal of rural settlements. The Yellow River Delta is a strategic region for ecological protection, agricultural production, and economic development in China. Under the combined influences of environmental protection policies, demographic shifts, and urban development, the spatial pattern of rural settlements in this region has undergone substantial changes in recent decades. Therefore, taking Dongying City in Shandong Province, the core area of the Yellow River Delta, as the study area, this research aims to quantify the impacts of rural settlement dynamics on regional carbon storage and explore potential carbon storage changes under an orderly rural settlement withdrawal scenario.On the basis of multisource remote sensing datasets and driving factor data, the InVEST (integrated valuation of ecosystem services and trade-offs) model was employed to estimate regional carbon storage and evaluate the impacts of rural settlement expansion on carbon storage during the study period. Subsequently, the PLUS (patch-generating land use simulation) model was coupled with the InVEST model to simulate future land-use patterns and assess carbon storage changes under an orderly rural settlement withdrawal scenario in 2040. The results are as follows: (1) From 1987 to 2023, despite the continuous decline in rural population, the total area of rural settlements in Dongying City exhibited an overall increasing trend. The cumulative expansion of rural settlements reached223.233 km², mainly encroaching upon other land-use types, particularly cropland. (2) During the same period, the expansion of rural settlements resulted in a reduction of approximately 0.555×10⁶ t in regional carbon storage. Among different land-use conversions, cropland occupation contributed the largest carbon storage loss, reaching 0.257×10⁶ t and accounting for 46.31% of the total carbon loss. (3) Simulation results indicate that under the orderly withdrawal scenario, the area of rural settlements could decrease by approximately 106.275 km² by 2040, leading to an increase of about 0.250×10⁶ t in regional carbon storage. The expansion of cropland would play a dominant role, contributing approximately 0.163×10⁶ t of carbon storage gain, accounting for 65.20% of the total carbon increment under this scenario.Overall, the orderly withdrawal of rural settlements can effectively promote the transformation of idle into ecological land or cropland, thereby enhancing regional carbon sequestration capacity.The findings of this study provide scientific support for ecological protection and food security strategies in the Yellow River Delta region.Keywords：carbon storage;rural settlement evolution;Yellow River delta;InVEST model;PLUS model137|36|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
City and Land
Atmospheric correction algorithm for HY-1 C/D images for remote sensing reflectance inversion of high-turbidity water AI Introduction
Abstract：The HY-1C/D satellites, equipped with the Chinese Ocean Color and Temperature Scanner (COCTS) and an ultraviolet imager (UVI), offer spectral capabilities—including ultraviolet (UV, 350—400 nm) and near-infrared (NIR, 745—885 nm) bands—that are widely used in traditional atmospheric correction for turbid waters. However, atmospheric correction over turbid waters (Case-Ⅱ water) remains challenging given limitations in traditional algorithms. Most existing methods rely on extrapolating aerosol scattering data from short-wave infrared (SWIR) or NIR reference bands, but errors in aerosol reflectance (ρα) inevitably amplify with increasing extrapolation distance. Additionally, strong absorption by aerosols in blue-violet bands (350—420 nm) introduces significant uncertainties in remote sensing reflectance (Rrs) retrievals, which cannot be resolved by single-reference-band approaches. This study aims to address these issues by proposing a new interpolation-based atmospheric correction algorithm for HY-1C/D satellite data, leveraging UV and NIR bands to improve accuracy across the entire spectrum, particularly in problematic shorter wavelengths.To thoroughly validate the algorithm’s accuracy, this study selected Hangzhou Bay—a representative highly turbid water body—as the test site. Given that the reference bands’ water-leaving radiance is not completely negligible, initial calibration of Rrs (UV) contributions was required. This study first developed a Rrs (UV) estimation model based on inherent optical properties (IOPs) parameters, utilizing a pregenerated simulated dataset of water spectral characteristics. Afterward, aerosol reflectance ρα (UV) for HY-1C/D satellite data was obtained under the “dark pixel” assumption. Subsequently, ρα (NIR) was derived by extrapolating from ρα (UV) using aerosol scattering spectral relationships. The image data were then classified on the basis of aerosol reflectance characteristics at both spectral ends (UV and NIR), and initial interpolation functions were constructed to produce preliminary correction results. The first-round interpolation results were incorporated into the Rrs (UV) estimation model to obtain higher-precision ρα (UV) values. The same procedure was repeated for a second interpolation, completing the entire atmospheric correction process.This study employed in-situ measurements from the “HTYZ” platform in Hangzhou Bay to validate the accuracy of the retrieval results. The single-day comparison revealed that after improving the calculation accuracy of ρα (UV), the second interpolation significantly enhanced retrieval precision. To further evaluate the algorithm’s practical performance, we conducted a comprehensive validation using long-term time-series data from January 2021 to December 2021. The results demonstrated that the relative mean deviation (RMD) across all spectral channels remained below 15%, confirming the algorithm’s applicability for HY-1C/D satellite data processing.This study developed an interpolation-based atmospheric correction algorithm specifically for highly turbid waters, using Hangzhou Bay as a representative case study, on the basis of data from the UVI and COCTS payloads aboard the HY-1C/D satellites. Comparative validation with long-term in-situ water spectral measurements demonstrated that the secondary interpolation correction algorithm achieved significantly improved retrieval accuracy after obtaining more precise ρα (UV) values. Owing to the absence of SWIR channels on the HY-1C/D satellites, the algorithm had to derive aerosol reflectance in the NIR channel from UV channel data to enable the interpolation-based retrieval. Looking ahead, while China’s next-generation ocean color satellite HY-1E, with its complete UV-SWIR spectral coverage, will eliminate the current limitation of missing SWIR reference bands, the methodology proposed in this study—which enhances atmospheric correction accuracy through Rrs (UV) contribution correction—remains valid and applicable.Keywords：HY-1C/D satellite;turbid water;atmospheric correction;remote sensing reflectance;IOPs74|42|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Fine-scale analysis of multitemporal SAR-Derived sea surface winds for Typhoon Yagi （2024） AI Introduction
Abstract：Typhoon Yagi (No. 202411) formed east of the Philippines and moved northwestward. After entering the South China Sea, it rapidly intensified and reached super typhoon status. The storm subsequently unleashed severe winds, torrential rain, and catastrophic flooding across the Pearl River Delta. A thorough characterization of the fine-scale surface wind structure and its temporal evolution is fundamental to unraveling the physical mechanisms governing typhoon intensity variations and to enhancing the skill of operational forecasts. Conventional observing systems are generally insufficient for resolving fine-scale surface winds, while synthetic aperture radar (SAR) provides a unique capability for all‑weather, high‑resolution wind retrieval. The growing constellation of SAR satellites now enables multi-source collaborative observations that are critical for resolving the fine‑scale wind‑field evolution and boundary‑layer structure of rapidly intensifying tropical cyclones.The present study exploits a multi‑temporal SAR dataset consisting of 13 images acquired by Sentinel‑1, Radarsat‑2, and RCM (all C‑band) throughout the lifecycle of Typhoon Yagi, of which 9 scenes covering the inner core were retained for analysis spanning from tropical storm to super typhoon stages. Using the 500‑m‑resolution sea surface wind retrieval products, we performed land masking and applied Gaussian smoothing to suppress speckle noise, along with a uniform filter to effectively remove scalloping and inter‑swath stitching artifacts inherent in wide‑swath SAR imagery. In addition, we developed a multi‑criterion automatic localization algorithm that identifies the typhoon center by detecting wind speed minima within the eyewall and accounting for the radial wind structure. Based on these processed wind fields, we extracted and analyzed the structural parameters during the evolution of Typhoon Yagi, including maximum wind speed, 95th and 99th percentile wind speeds, 50‑kt and 64‑kt wind radii, and the radius of maximum wind, and further applied azimuthal Fourier decomposition to quantify the wind field asymmetry.The results reveal that during rapid intensification, the radius of maximum wind contracted markedly while maximum wind speed, 50 kt and 64 kt wind radii expanded considerably, and the winds became more asymmetric than before, all of which stabilized upon the storm’s attainment of super typhoon intensity. Owing to its 500‑m spatial resolution, the SAR‑derived wind field allowed for the unambiguous identification of kilometer‑scale roll vortices in the atmospheric boundary layer. Two‑dimensional Fourier transform spectral analysis was employed to extract the dominant orientations and wavelengths of the roll vortices. revealing that these wavelengths are predominantly concentrated between 2 and 3 km, and that the roll orientations are broadly aligned with the ambient shear field of the typhoon. Furthermore, the spatial distribution of the rolls shows no significant variation between the severe typhoon and super typhoon stages, which is consistent with the prevailing understanding that the formation and maintenance of roll vortices are primarily governed by local wind‑shear instability dynamics.Despite limited sample size, this study demonstrates the unique potential of high-resolution SAR in capturing both axisymmetric structural evolution and kilometer-scale boundary-layer processes of typhoons. The established preprocessing and center-localization methodologies provide a robust framework for future large-sample multi-typhoon analyses, while underscoring the urgent need to incorporate boundary-layer roll effects into numerical parameterization schemes, given that current coarse-resolution models fail to resolve these fine-scale dynamical features essential to air-sea interaction.Keywords：synthetic aperture radar (SAR);Typhoon Yagi;Wind structure;Wind radius;Asymmetry;Boundary-layer rolls311|2269|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21CNN-GRU-based retrieval method for significant wave height under high sea states using spaceborne GNSS-R Data AI Introduction
Abstract：Significant Wave Height (SWH) is a key parameter in wave monitoring, which is of great significance for the study of ocean climate change and marine disaster early warning. As a new operational remote sensing technology, spaceborne GNSS-R has been widely applied to the observation of SWH. Existing studies have shown that SWH retrieval has achieved high accuracy under low and medium sea states. However, data saturation and inadequate feature extraction under high sea states limit the further improvement of retrieval accuracy in such conditions. This study proposes an SWH retrieval method for high sea states based on CYGNSS data, providing a reference for the wide application of CYGNSS data in typhoon-affected areas.In this paper, a CNN-GRU-based SWH retrieval model for high sea states is proposed; it combines CNN and GRU to achieve the synergistic extraction of spatial and temporal features from GNSS-R data. The CNN branch processes the DDM to extract spatial features and introduces a spatial attention mechanism to dynamically enhance the feature extraction capability near the specular reflection points, while the GRU branch processes one-dimensional parameters to extract the time-series features of the GNSS-R observations. The features extracted by the two branches are fused by an FCN to produce the final SWH values.Experimental results show that the model performance is optimal when wind speed and rainfall data are added to the CNN-GRU model, with an RMSE of 0.55 m, an MAE of 0.41 m, and a PCC of 0.91. Ablation experiments verify that wind speed contributes more to improving the SWH retrieval accuracy for high sea states. To prove the stability and cross-sea universality of the CNN-GRU model, this study utilizes the data of North Atlantic hurricanes. The results show that the performance of the CNN-GRU model is better than that of the existing SWH retrieval model.This study proposes a CNN-GRU-based GNSS-R retrieval method for SWH under high sea states, using CYGNSS data as the primary data source. The results of this study show that CNN-GRU is feasible and effective in retrieving SWH for high sea states. Future work should include validating the model with additional in-situ measurement data.Keywords：GNSS-R;Significant wave height;high sea states;CYGNSS;deep learning;DDM70|30|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Atmosphere and Ocean
Abstract：The Surface Water and Ocean Topography (SWOT) mission represents a major advancement in satellite altimetry for inland water monitoring given its wide-swath interferometric capability. Although SWOT significantly improves spatial coverage and resolution compared with conventional nadir altimeters, inland observations remain affected by strong random noise and residual systematic errors, especially in regions far from ocean cross-calibration areas. These errors limit the accuracy and reliability of water level estimation over lakes, reservoirs, and rivers. This study aims to systematically analyze the characteristics of random and residual systematic errors in SWOT L2_HR_PIXC products and to develop an effective error compensation and correction framework for inland water level observations.Taihu Lake, Chaohu Lake, Qinghai Lake, and multiple reservoirs and rivers in Liaoning Province and the Taihu region were selected as study sites. A two-step denoising strategy was first developed to mitigate large abnormal outliers caused by random errors. Coarse denoising constrained observations within reasonable elevation intervals using water masks and adaptive histogram filtering, while fine denoising employed an improved density-based clustering algorithm and RANSAC-based robust fitting to remove local outliers. After noise suppression, the spatial relationship between water surface elevation and cross-track distance was analyzed for large lakes to reveal the structure of residual systematic errors. On the basis of the assumption that lake surfaces are approximately horizontal, a residual systematic error correction model was constructed. The first-order slope term was derived from multi-lake cross-track fitting, and the constant term was determined using high-precision water level measurements from a control station.After denoising, SWOT water level observations achieved decimeter-level accuracy, with RMSE values ranging from 0.1 m to 0.5 m across different stations. The residual systematic error exhibited distinct spatial patterns. It was approximately linearly correlated with cross-track distance in near-sea regions, whereas in far-sea areas, a quadratic trend was observed. The proposed correction approach effectively reduced these errors. In Liaoning Province, the overall RMSE decreased by 0.07 m (from 0.30 m to 0.23 m). In the Taihu Lake region, characterized by a greater distance from the ocean, the RMSE decreased by more than 0.30 m and was reduced to 0.16 m after correction. Even in extreme cases lacking cross-calibration information, the proposed method effectively constrained meter-level errors to the decimeter scale, highlighting its strong robustness and reliability.This study establishes a systematic processing framework for SWOT inland water level refinement, including robust denoising and residual systematic error compensation. The proposed method requires a minimum of one control station's measurement data and can significantly improve inland water level accuracy without replacing the official cross-calibration procedure. Results demonstrate that residual systematic errors over inland regions exhibit predictable cross-track dependence and can be effectively modeled and corrected. This study provides theoretical insight and practical guidance for high-precision application of SWOT data in inland hydrological monitoring.Keywords：inland water bodies;satellite altimetry;water level monitoring;SWOT;random error;two-step denoising method;system error;error correction67|35|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Spatiotemporal fusion algorithm for remote sensing images based on the combination of GAN and NSST AI Introduction
Abstract：Currently, most deep learning-based methods for the spatiotemporal fusion of remote sensing images employ convolutional operations to extract image features, which are then combined to produce higher-accuracy fused images. However, because of the convolution operation balancing pixel errors within local regions during the process of extracting image features, the high-precision fused image obtained after integrating various extracted features still suffers from issues such as weak texture details and blurred region boundaries.To address such issues, in this paper, we propose a new spatiotemporal fusion method based on the combination of Generative Adversarial Network (GAN) and nonsubsampled Shearlet transform (NSST). This method is based on the GAN framework and consists of a generator and a discriminator. Specifically, a dilated convolution branch and a subtraction-attention module (SUB-AM) are first introduced into the encoder structure of the generator, so that the extracted fine features can take into account the global nature; the coarse feature changes captured by the SUB-AM block are used to adjust the extracted fine features to improve the prediction accuracy. Then, through NSST and an adaptive reconstruction module (NSST-ARM), the high- and low-frequency information of the results generated by the encoder-decoder and the fine image of the reference moment are reconstructed in accordance with adaptive fusion rules, enhancing the high-frequency details of the fused image. Finally, a clarity loss term is introduced into the loss function to reduce the clarity loss of the fused image. The clarity loss term consists of the average gradient loss and the spatial frequency loss.Experimental results show that compared with a state-of-the-art method, the proposed method exhibits increased structural similarity value by 1.5% on average, reduced relative dimensionless global error in synthesis value by 14% on average, and improved texture detail expression and edge structure clarity of the fused image. Experiments prove the effectiveness of the proposed method.In this paper, we propose a spatiotemporal fusion method for remote sensing images based on the combination of GAN and NSST. The method has the following advantages: (1) Dilated convolution and SUB-AM are introduced into the encoder to make the features of the fused image accurate and the image information comprehensive. (2) The introduction of NSST-ARM and the optimization of the loss function make the edges and contours of the fused image clear. Experimental results on three datasets show that the method proposed in this paper can achieve high fusion accuracy and obtain clear image details. Nonetheless, the applicability of the proposed method remains limited; that is, it cannot be applied to all types of land. Moreover, there is still considerable room for improvement in its efficiency. Given these limitations, how to make our method more generalizable and efficient is a topic worth exploring in future research.Keywords：spatiotemporal fusion;Generative Adversarial Network (GAN);dilated convolution;Nonsubsampled Shearlet Transform(NSST)97|38|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Spatiotemporal change collaborative perception method for remote sensing semantic change detection AI Introduction
Abstract：Semantic Change Detection (SCD) aims to accurately identify complex land cover changes and simultaneously determine their corresponding semantic categories from bitemporal remote sensing images. Existing methods typically employ separate change localization and semantic recognition branches to accomplish SCD tasks. However, these frameworks often fail to fully integrate and leverage semantic information to guide the change detection process, which reduces their robustness in complex scenarios, seasonal and illumination variations, and multiscale small-target changes. To address this limitation, this study proposes a Semantic-guided Spatio-Temporal Collaborative Perception Network (SemSTNet) that incorporates semantic guidance into the precise identification of change regions while considering spatiotemporal interactions of bitemporal features.The proposed SemSTNet consists of three key components. First, a symmetric spatiotemporal demodulation differential module (SSDM) is constructed. It employs bidirectional feature modulation to achieve cross-temporal feature mutual calibration and explicitly model intrinsic spatiotemporal difference patterns using 3D central difference convolution, thereby extracting discriminative difference features from bitemporal images. Second, a Semantic Gated Fusion Decoder (SGFD), which introduces a semantic similarity dynamic weighting mechanism, is designed. It generates semantic priors through cosine similarity maps and enhances the saliency representation of genuine change regions via adaptive gated fusion. Third, a Contrastive Change Loss (CCL) is proposed to construct dual supervision signals. It constrains the semantic consistency of unchanged regions in the feature space and enlarges the interclass distance between changed and unchanged regions through margin-based contrastive learning.Extensive experiments are conducted on two SCD datasets: the SECOND dataset and the JL1 dataset. On the SECOND dataset, SemSTNet achieves the best performance with an mIoU of 73.62%, a separated kappa (SeK) of 24.19%, and an F1 score of 64.27%, surpassing the second-best method CdSC by 0.30%, 0.67%, and 0.57%, respectively, while maintaining lower parameter count (27.85 M vs. 38.6 M) and reduced memory consumption (424.17 MB vs. 482.9 MB). On the JL1 dataset, SemSTNet again achieves the highest scores across all metrics, with an mIoU of 86.51%, a SeK of 59.02%, and an F1 of 88.01%, demonstrating strong robustness against complex seasonal spectral variations. Ablation studies on the SECOND dataset confirm the effectiveness of each proposed component, with the full model improving mIoU from 70.60% to 73.62% and SeK from 19.09% to 24.19% over the baseline.The proposed SemSTNet effectively breaks the conventional separation between change detection and semantic recognition by establishing a collaborative optimization mechanism among SSDM, SGFD, and CCL. Experimental results on multiple public datasets validate the superiority of the proposed method in change localization accuracy and semantic recognition consistency, particularly in scenarios with frequent land cover type transitions and significant scale variations. Future work will focus on exploring lightweight collaborative architectures and incorporating multimodal remote sensing data to further enhance practical applicability.Keywords：remote sensing images;land cover change;semantic change detection;spatio-temporal collaboration;semantic guidance91|40|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Machine learning-driven remote sensing inversion model for suspended sediment concentration and sensitivity analysis AI Introduction
Abstract：Remote sensing inversion offers an efficient approach for monitoring Suspended Sediment Concentration (SSC). However, its applicability for rivers with high and wide-range SSC requires further validation. This study aims to develop a machine learning model based on recursive feature elimination with cross-validation and random forest (RFECV-RF) to retrieve SSC from multispectral imagery, using the Shizuishan and Wubu stations on the main stream of the Yellow River as study sites.Sentinel-2 multispectral reflectance data were matched with in-situ daily SSC measurements from 2020 to 2024. After outlier removal using the isolation forest algorithm, datasets of 267 and 256 samples were obtained for the Shizuishan and Wubu Station, respectively. Spectral features, including single bands and band combinations (ratios and differences) from visible to shortwave infrared, were constructed. The RFECV method was applied to select optimal features, and a Random Forest model was subsequently built to establish the relationship between spectral features and SSC. Model performance was evaluated using R2, RMSE, and MAE. Spectral sensitivity was analyzed using regression fitting and slope decay ratios. Parameter sensitivity was assessed via the perturbation method by varying key parameters and evaluating corresponding changes in performance metrics using sensitivity coefficients. Uncertainty arising from data input, model parameters, and model structure was quantified using a Monte Carlo simulation framework with 500 iterations, and variance decomposition was applied to quantify the contribution of each uncertainty source.The RFECV-RF model achieved R2 values greater than 0.8 for both stations, with RMSE below 1.6 kg/m3, demonstrating reliable SSC estimation within the range of 0—44.5 kg/m3, though with underestimation of extreme high SSC values. A strong nonlinear relationship was observed between spectral features and SSC. The B8A band, along with B7/B5 and B8-B11, remained sensitive across the entire SSC range, while features such as B3/B8, B4/B8, and B5/B6 tended to saturate as SSC increased. Parameter sensitivity analysis indicated that model performance was robust, with the most sensitive parameters being the subsampling rate and the number of features removed per iteration. Uncertainty analysis revealed that data input and model structure were the dominant sources of uncertainty, contributing over 98% of total variance, with larger uncertainties observed for low SSC values when modeling wide-range SSC. When applied to river reaches, the inversion model effectively captured temporal variations and spatial transport dynamics of SSC, with higher SSC values in the middle reaches than in the upper reaches, consistent with actual observations.The RFECV-RF model provides a reliable and effective approach for retrieving high and wide-range SSC in sediment-laden rivers using multispectral remote sensing. The identified key spectral features and quantified uncertainties offer valuable insights for improving inversion accuracy and support the development of automated SSC monitoring methods for complex river systems.Keywords：SSC;remote sensing inversion;RFECV;RF;Yellow River;spectral sensitivity;uncertainty205|106|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21High-resolution optical remote sensing image change detection based on Cross-Transformer difference feature fusion AI Introduction
Abstract：Deep learning-based Change Detection (CD) methods, especially Convolutional Neural Networks (CNNs), are now widely used for high-resolution optical remote sensing imagery. Nevertheless, CNN-based architectures retain important structural limitations. Their restricted receptive fields make it difficult to capture global contextual cues that are essential for modeling long-range spatial dependencies and interpreting complex geographic scenes. In addition, insufficient interaction among multi-level features weakens the integration of high-level semantic representations with low-level spatial details, thereby limiting the discrimination of subtle and irregular changes.To overcome these limitations, this paper presents the Cross-Transformer-based Difference Feature Fusion Network (CTDFFNet), a unified CD framework that combines global context perception, cross-level feature interaction, and adaptive feature refinement. CTDFFNet follows an encoder-decoder architecture. In the encoder, a ResNet18 backbone pre-trained on ImageNet extracts multi-scale local feature representations from bi-temporal remote sensing images. To compensate for the limited receptive field of standard convolutions, a Cross-Transformer module is designed to learn discriminative difference features. Through a hierarchical cross-attention mechanism, the module captures long-range spatial dependencies and strengthens interactions across feature levels, thereby improving sensitivity to changes under complex conditions, including illumination variation, shadow effects, and seasonal differences. In the decoder, dense connections enable each layer to receive information from all preceding layers, which promotes feature reuse, improves the use of hierarchical information, alleviates gradient vanishing, and supports more stable training. An Adaptive Channel Enhancement Module (ACEM) is further incorporated to dynamically recalibrate channel-wise feature responses by learning attention weights, emphasizing change-salient features while suppressing irrelevant background noise. A Sigmoid layer is then used to produce the final change map.Extensive experiments on three publicly available CD datasets, namely LEVIR-CD, CDD, and WHU-CD, were conducted to evaluate the effectiveness and generalization capability of CTDFFNet. Quantitative results demonstrate that CTDFFNet consistently outperforms state-of-the-art methods on all three datasets, achieving the highest F1 scores of 91.52%, 95.88%, and 93.02% on LEVIR-CD, CDD, and WHU-CD, respectively. These gains confirm the effectiveness of incorporating Transformer-based mechanisms into CD frameworks. Visual comparisons further indicate that CTDFFNet effectively suppresses interference caused by complex backgrounds and environmental variations, producing change maps with substantially fewer false alarms and missed detections while maintaining high precision in delineating changed objects at different scales.In summary, CTDFFNet addresses key limitations of CNN-based CD frameworks through three main innovations: (1) the Cross-Transformer module captures long-range dependencies and promotes cross-level feature interaction, enhancing sensitivity to complex scene changes; (2) the densely connected decoder improves multi-scale feature use and alleviates gradient vanishing, contributing to stable training; and (3) ACEM dynamically emphasizes change-salient features and suppresses background interference. Experiments on multiple datasets demonstrate that CTDFFNet outperforms existing methods and exhibits strong generalization and robustness under challenging conditions. These findings support its potential for a wide range of remote sensing CD applications. Future work will extend the framework to multi-modal CD scenarios and investigate label-efficient learning strategies to reduce dependence on large-scale annotated data.Keywords：change detection;deep learning;Cross-Transformer;cross attention mechanism;adaptive channel enhancement103|44|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Distillation method for remote sensing image semantic segmentation via frequency-domain filtering AI Introduction
Abstract：With the increasing deployment of remote sensing semantic segmentation models on edge devices, achieving an appropriate balance between segmentation accuracy and inference efficiency under limited computational resources has become a critical challenge. This study addresses high-frequency noise interference in front-end encoder features during knowledge distillation and aims to design a reliable and efficient distillation framework for remote sensing semantic segmentation in resource-constrained edge computing environments.This study proposes a Frequency Domain Filtering Distillation (FDF) framework. In the Fourier domain, a central square mask is applied to the teacher and student features, enabling selective transfer of structurally meaningful low-frequency semantic knowledge while suppressing irrelevant high-frequency noise components. For the distillation architecture, U-Lite, a lightweight network with compact parameters and fast inference that has been validated in medical imaging tasks, is used as the student model to ensure edge-device efficiency. Comprehensive experiments on three widely used remote sensing semantic segmentation datasets (LoveDA, Vaihingen, Potsdam) are conducted to evaluate the proposed method.Experimental results indicate that FDF provides competitive performance gains and stronger training and inference stability for remote sensing semantic segmentation. Specifically, FDF substantially reduces the risk of performance degradation: the standard deviation of ΔmIoU is only 1.03, which is lower than that of CWD (1.25) and Fourier (1.34), and it avoids the maximum 2.37% performance drop observed in other distillation methods. In terms of efficiency, the FDF-based model requires only 3.3G FLOPs of computational overhead and reaches an inference speed of 16.63 ms/frame, which is about 2.5 times faster than the IFVD and Fourier methods.The proposed FDF framework effectively mitigates high-frequency noise interference during feature distillation for remote sensing semantic segmentation models, and its integration with the lightweight U-Lite model further ensures model efficiency. By balancing segmentation accuracy, model stability, and inference efficiency, FDF provides an effective, reliable, and efficient technical solution for deploying remote sensing semantic segmentation in resource-constrained edge computing environments.Keywords：deep learning network;distillation of knowledge;semantic segmentation;feature extraction;fast Fourier transform76|28|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Spatiotemporal-spectral random forest reconstruction of orbital gaps in AMSR2 brightness temperature AI Introduction
Abstract：Brightness Temperature (BT) serves as a critical geophysical parameter reflecting the Earth’s surface energy budget and holds substantial significance in meteorological monitoring, climate change studies, and environmental remote sensing. The Advanced Microwave Scanning Radiometer 2 (AMSR2) enables global, all-weather acquisition of BT data. However, because of orbital design constraints and sensor-specific characteristics, the AMSR2 BT product suffers from orbital data gaps, which hinder its continuity and limit its effectiveness in quantitative applications. This study aims to develop a robust reconstruction framework to address these data discontinuities and improve the spatiotemporal integrity of AMSR2 BT datasets.To reconstruct the missing orbital BT data, we propose a novel multidimensional model termed spatiotemporal-spectral random forest (STSRF). The model integrates spatiotemporal patterns with generalized spectral features that characterize the nonlinear relationships between BT and multiple environmental variables. Using the 2020 AMSR2 BT dataset over China as a case study, we separately constructed training and validation datasets on a monthly basis for daytime and nighttime observations. Simulated data masking and real-missing data scenarios were applied to evaluate reconstruction accuracy. Moreover, downscaling experiments were conducted before and after reconstruction to assess the influence of data integrity on spatial detail restoration.In the simulation experiments, the STSRF model demonstrated high reconstruction accuracy, with root-mean-square errors ranging from 0.78 K to 2.21 K during the day and from 0.73 K to 2.35 K at night. Notably, the model achieved superior performance in high-altitude regions. In real experiments, the spatial distribution of reconstructed BT closely matched original AMSR2 observations, showing no discernible reconstruction artifacts. Downscaling evaluation revealed that the STSRF-based reconstruction significantly enhanced numerical precision and spatial feature retention compared with the direct downscaling applied to incomplete datasets. The enhanced numerical precision and spatial feature retention observed in the downscaling evaluation confirm the added value of gap filling prior to resolution enhancement.The proposed STSRF model effectively overcomes the limitations of traditional single-feature approaches by incorporating multidimensional information, including temporal dynamics and environmental variability. It provides a reliable strategy for reconstructing AMSR2 BT orbital gaps while enhancing downstream data usability for high-resolution microwave remote sensing applications. The results affirm that spatiotemporal and spectral feature fusion not only improves reconstruction robustness but also supports accurate and complete environmental monitoring over large regions.Keywords：AMSR2 brightness temperature;orbital gap reconstruction;spatiotemporal spectral random forest;downscaling;environmental variables80|34|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21Lightweight remote sensing image target detection based on spatial-channel reconstruction AI Introduction
Abstract：In the field of computer vision, lightweight target detection in remote sensing images is an academically challenging research topic. With the continuous improvement in the accuracy of remote sensing target detection, the complexity of models has increased significantly. Existing remote sensing object detection methods struggle to strike a good balance between accuracy and model complexity. Unlike conventional natural images, remote sensing images have a large imaging area, limited available information, and a multitude of target types. The vast amount of background information can easily interfere with detection, and the excessive number of target categories makes it difficult to achieve high-precision detection of specific targets. Moreover, targets in remote sensing images are densely packed and arbitrarily rotated and exhibit significant scale variations, posing a great challenge for target localization. While the accuracy continues to improve in remote sensing image target detection, models have become increasingly complex in pursuit of high precision. These complex models’ heavy computational load and large parameter count severely restrict practical application on resource-constrained devices; model lightweighting has become a key factor in promoting technological application.To address the aforementioned issues, this study proposes a lightweight remote sensing image target detection method based on spatial-channel reconstruction, aiming to maintain high-precision detection of circular and rectangular targets while reducing model complexity. First, in response to the redundant information within feature maps at the spatial dimension, a spatial reconstruction unit is adopted. It separates feature maps on the basis of the richness of spatial features (i.e., into groups with abundant spatial information and groups with redundant spatial information) and then performs cross-reconstruction operations on the two groups to reduce spatial feature redundancy and enhance the spatial feature representation of remote sensing image targets. Subsequently, to address the redundancy of channel information in feature maps, a partial convolution-based channel reconstruction unit is proposed. It divides feature maps into two parts along the channel dimension: One part uses partial convolution for efficient feature extraction, and the other part uses pointwise convolution to obtain hidden detailed information. The two parts are weighted, reconstructed, and then concatenated to extract features at a decreased computational cost and enhance the representation of important channels.The lightweight method based on spatial-channel reconstruction is applied to ACRFNet, an adaptive circular receptive field network for remote sensing image target detection. Experimental results on the NWPU VHR-10, DIOR, and HRRSD datasets verify its performance. Specifically, the method can maintain or even improve detection accuracy while significantly reducing the model’s computational load and parameter volume. Compared with other mainstream methods for remote sensing image target detection, the proposed method demonstrates obvious advantages in balancing detection accuracy and model complexity, especially for circular and rectangular targets that are difficult to detect.The proposed method effectively solves the core problem of balancing detection accuracy and model complexity. By reducing feature redundancy, it overcomes the deployment limitations of complex models on resource-constrained devices, retains good adaptability to typical targets, and provides reliable technical support for the practical application of remote sensing target detection technology.Keywords：remote sensing imagery;object detection;circular target;model lightweighting;reconstruction unit261|263|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-07-21
Models and Methods
