Issue Ebookcover storyIssue CoverIssue Contents
Latest Issue
Online FirstList of IssuesDocument Types
2026
Issue 8
Issue 7
Issue 6
Issue 5
Issue 4
Issue 3
Issue 2
Issue 1
2025
2024
2023
2022
2021
2020
2019
2018
2017
Abstract：The changes the Earth’s radiation budget directly impact the stability of the climate system and its future evolution trends. Accordingly, accurately quantifying the components of the radiation budget is of great theoretical and practical significance for deepening our understanding of climate change, improving climate change prediction and response capabilities, and promoting its applications in various fields such as land, atmosphere, and oceans. As an indispensable approach for Earth system radiation budget monitoring, remote sensing has achieved remarkable advancements in radiation budget observation over recent decades. Despite these fruitful achievements, the quantitative retrieval of Earth’s radiation budget remains confronted with substantial challenges stemming from the complex nonlinear coupling effects among atmospheric, cloud, and land surface factors. This paper systematically reviews the state-of-the-art progress in remote sensing-based radiation budget monitoring and provides prospective insights for its future development.This paper elaborates on research advances from two perspectives. On one hand, it systematically reviews the latest progress in the field of radiation budget from space, alongside the core achievements obtained by the authors based on successive funded research projects, particularly in the areas of all-weather (clear and cloudy skies), all-time (daytime and nighttime), all-wave (longwave and shortwave), multi-components (surface and TOA), and multi-scenario (flat surface, rugged terrain and polar regions) radiation components derivation algorithms and scientific product development. Major advances are summarized as follows: (1) For TOA radiation, a universal multi-band Lookup Table (LUT) approach is established for TOA albedo retrieval, suitable for both broadband and narrowband sensors. Furthermore, we present a universal multidimensional matrix mapping (MDMAP) algorithm to bridge narrowband and broadband sensors and support efficient outgoing longwave radiation estimation; (2) For surface radiation, a suite of innovative frameworks is proposed to simultaneously retrieve multiple radiation fluxes under all-sky conditions; (3) Regarding sub-cloud longwave radiation reconstruction, a Multivariate Data Interpolating Empirical Orthogonal Functions (MDINEOF) method is adopted to dynamically reconstruct sub-cloud land surface temperatures with the constraint of radiation fluxes; (4) For temporal upscaling, a Diurnal Variation Based (DVB) scheme is described. Combined with ERA5 reanalysis diurnal prior information, this scheme converts instantaneous satellite observations into temporally continuous radiation; (5) To mitigate prominent topographic effects over mountains, an Uniform Shortwave Topographic Radiation Model (USWTRM) is introduced for accurate estimation of shortwave radiation over rugged terrain. Results demonstrate that these advanced algorithms substantially enhance the spatiotemporal continuity, physical consistency and accuracy of Earth radiation budget derivations.On the other hand, based on state-of-the-art research findings and international frontiers of radiation budget studies, this paper outlines key research priorities for future Earth radiation budget remote sensing: (1) Development of novel Earth observation techniques and monitoring spectra for radiation flux measurements; (2) Integration of AI and physical algorithms enables multi-component, synergistic optimization and inversion;(3) Radiation estimation over complex and extreme environments, including mountainous areas, polar regions, fog-prone zones and regions with strong convective activities; (4) Development of radiation products featuring spatiotemporal continuity, as well as complete coverage of both short- and long-wave radiation components, within a framework of multi-source data fusion frameworks; (5)Verification and uncertainty quantification for remote sensing inversion algorithms and products.This work seeks to promote theoretical and methodological advances in quantitative remote sensing. Specifically, it delivers insights into quantitative remote sensing for complex environments and furnishes scientific evidence and data support to deepen global change research and meet relevant national strategic demands.Keywords：radiation budget;outgoing longwave radiation;shortwave radiation;longwave radiation;rugged terrain;albedo;land surface temperature10|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Optical remote sensing provides rich spectral information and high spatial resolution for fine-scale urban monitoring, but is limited by weather and illumination conditions. Conversely, SAR offers all-weather, all-day observation capabilities, although its interpretation is hindered by geometric distortions and speckle noise. Given their complementary characteristics, optical-SAR integration, empowered by recent advances in Artificial Intelligence (AI), has emerged as a promising approach for urban monitoring and change detection. This paper reviews the state of the art in AI-enabled optical-SAR collaborative urban monitoring and identifies key challenges and future opportunities.A systematic literature review was conducted based on more than 900 publications indexed in the Web of Science Core Collection and the China National Knowledge Infrastructure (CNKI) between 2001 and 2025. The reviewed studies were analyzed from the perspectives of major urban monitoring targets and dominant methodological frameworks. Particular attention was paid to the role of AI in enhancing multimodal data fusion and temporal analysis, thereby enabling more effective integration of optical and SAR observations. The review also examined key technical components and challenges associated with optical-SAR synergy, including precise cross-modal co-registration, modality information gaps, spatiotemporal scale effects, the construction of high-quality training datasets, and the interpretability of AI-driven models.The analysis reveals that recent advances in AI have substantially improved the ability to integrate optical and SAR observations, leading to significant progress in urban monitoring, land-cover mapping, and change detection. Optical-SAR fusion has emerged as an effective solution to overcome the limitations of individual sensors and improve the robustness, accuracy, and temporal continuity of urban observations. The review further identifies several critical challenges that continue to limit the operational deployment of optical-SAR collaborative monitoring, including cross-modal geometric inconsistencies, heterogeneous feature representations, scale mismatches, insufficient benchmark datasets, and the limited interpretability of deep learning models. At the same time, a number of emerging opportunities have been observed, including the development of advanced sensors and computing constellations, the widespread adoption of cloud computing platforms, the rapid evolution of large AI models, and increasing demand for Sustainable Development Goals (SDGs) related applications.Optical-SAR collaboration provides a key pathway toward high-precision, all-weather urban monitoring. AI-driven multimodal analysis is rapidly transforming the field and expanding its applications in urban planning, management, and sustainability assessment. Future research should focus on cross-modal representation learning, interpretable and generalizable AI models, large-scale benchmark datasets, and the integration of emerging technologies, including large AI models, cloud computing platforms, and next-generation satellite systems.Keywords：optical remote sensing;SAR;multimodal remote sensing;urban monitoring;AI;change detection;deep learning9|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Reviews
Abstract：In complex urban environments, dense buildings and significant height variations amplify nonlinear radiometric differences and geometric distortions among multimodal remote sensing images. These challenges severely hinder high-precision registration of high-resolution imagery. Existing registration approaches predominantly rely on two-dimensional features such as texture, often neglecting the valuable three-dimensional spatial information inherent in such scenes. To address these limitations, this study proposes a novel multimodal high-resolution remote sensing image registration method that integrates modality transformation with three-dimensional spatial relationship constraints, aiming to improve robustness and accuracy in complex urban scenarios.The proposed method consists of three main components. First, a cross-modal image translation technique is employed to reduce radiometric discrepancies between multimodal images, effectively narrowing the modality gap and facilitating subsequent feature extraction. Second, monocular depth estimation is introduced to efficiently generate depth maps from single images. These depth maps provide essential spatial priors and serve as a foundation for constructing more informative feature descriptors and enforcing spatial constraints. Finally, a matching strategy based on depth-guided 3D spatial relationship constraints is developed. This strategy includes multi-feature map keypoint detection to capture potential salient features, the construction of depth-enhanced joint descriptors to improve feature distinctiveness, and the incorporation of 3D spatial relationship constraints to ensure geometrically consistent matching. Together, these steps enable reliable detection, robust description, and accurate matching of feature points across multimodal images.The proposed method was compared with several traditional and deep learning methods on four representative multimodal remote sensing datasets. Experimental results show that the proposed method achieves an average Number of Correct Matches (NCM) of 510, which improves upon existing methods by a factor of 0.92 to 5.22. The average Root Mean Square Error (RMSE) is 1.58 pixels, and the registration accuracy is improved by a factor of 4.12 to 4.78 compared to state‑of‑the‑art methods. These results demonstrate that the proposed method has clear advantages in registration accuracy and overall performance, effectively suppressing nonlinear radiometric differences and geometric distortions in urban multimodal images and achieving high‑precision registration.This study presents a robust solution for multimodal high-resolution remote sensing image registration in complex urban environments. By combining cross-modal translation, monocular depth estimation, and depth-based three-dimensional spatial relationship constraints, the proposed method successfully addresses both nonlinear radiometric differences and geometric distortions. The integration of 3D spatial information significantly improves feature matching robustness and registration precision compared to conventional 2D-based approaches. Experimental validation confirms that the method achieves superior performance in terms of both accuracy and stability, establishing a solid foundation for downstream applications such as urban mapping, change detection, and multi-sensor data fusion.Keywords：image registration;multimodal imagery;3D spatial relationships;complex urban scenes;depth maps480|160|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Optical remote sensing images are commonly contaminated by clouds. Although deep learning-based methods have received considerable attention in cloud detection tasks, their ability to identify thin clouds is limited due to their sparse distribution, semi-transparency, and similarity to the background areas of thin clouds. Thus, effectively distinguishing thin clouds from thick clouds is challenging, and few studies have focused on this issue. This paper considers cloud detection as a semantic segmentation problem and proposes a transformer-based model for joint detection of thick and thin clouds (TTFormer) to simultaneously detect thick and thin cloud contamination in remote sensing images.This model enhances multiscale feature expression ability by fusing the local-global feature modulation module to effectively capture the dense texture of thick clouds and the semi-transparent characteristics of thin clouds. It adopts conditional position encoding to embed spatial information, addressing the challenges posed by the blurred boundaries of thin clouds. Moreover, it combines an edge perception branch to strengthen the retention of cloud boundary details.The performance of the TTFormer model was experimentally validated on the Landsat 8 Biome and CloudSEN12 datasets. Results show that TTFormer can more accurately distinguish thick clouds, thin clouds, and shadows under different land cover backgrounds. It is more accurate than existing six cloud detection methods, with an overall accuracy reaching 92.27% and 89.28% on the Landsat 8 Biome and CloudSEN12 datasets, respectively.The proposed TTFormer can effectively distinguish thin clouds from thick clouds, and it consistently demonstrates greater accuracy and stability under various land cover backgrounds.Keywords：cloud detection;image segmentation;Transformer;Landsat7|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Remote sensing image scene classification plays an important role in natural resource monitoring, territorial spatial planning, disaster emergency response, and other Earth observation applications. However, the performance of existing classification methods generally depends on large-scale, high-quality annotated datasets. The acquisition and manual annotation of remote sensing images are time-consuming and expensive, hence resulting in a pronounced contradiction between the strong demand for labeled data and the limited availability of annotations. Few-shot learning provides a promising solution by enabling models to generalize to novel categories by using only a small number of labeled samples. However, most conventional few-shot classification methods rely primarily on visual features and are therefore prone to confusing scene categories that exhibit similar visual appearances but differ in semantic meaning. Although recent vision language models can introduce textual priors to alleviate semantic ambiguity, they often treat all feature channels equally and overlook differences in the discriminative information carried by individual dimensions. This study aims to improve the discriminative capability and generalization performance of few-shot remote sensing scene classification by effectively integrating visual representations, textual semantic priors, and channel-level feature selection.A hybrid metric learning method based on Contrastive Language-Image Pre-training (CLIP) is proposed for few-shot remote sensing image scene classification. First, the image encoder and text encoder of CLIP are employed to extract visual representations of remote sensing images and semantic representations of category-level textual prototypes, respectively. These multimodal features are mapped into a unified embedding space to establish image-text semantic alignment. Second, a discriminative feature selection strategy jointly constrained by the mean inter-class similarity and inter-class variance is designed. The strategy evaluates the discriminative contribution of each feature channel by simultaneously considering the average similarity between different categories and the dispersion of inter-class relationships. Highly discriminative dimensions are retained, whereas redundant and weakly informative channels are suppressed, resulting in a more compact and separable feature space. Finally, a tripartite relational modeling module is constructed among query images, support images, and category textual prototypes. By jointly modeling image-to-image visual relationships and image-to-text semantic relationships, the module integrates sample-level visual evidence with category-level semantic knowledge and generates more reliable classification decisions through hybrid metric fusion.Extensive experiments were conducted on three fine-grained remote sensing scene classification datasets, namely, MEET, AID, and NWPU. The proposed method was compared with multiple representative few-shot learning methods and vision-language-based baseline approaches under different few-shot settings. The experimental results demonstrate that the proposed method achieves consistently superior classification performance across the three datasets, particularly under severely limited annotation conditions. The ablation results further verify the effectiveness of the proposed method and its individual components. In addition, the results indicate that discriminative feature selection can reduce interference from redundant channels and enhance inter-class separability, while category textual prototypes can alleviate semantic confusion among visually similar scene categories and improve the stability of classification decisions.The proposed CLIP-based hybrid metric learning method effectively combines image-text aligned representations, discriminative channel selection, and tripartite cross-modal relational modeling. By exploiting both visual information and textual semantic priors, the method improves the compactness of intra-class representations, the separability of inter-class features, and the generalization capability of the classifier. The proposed framework provides an effective and feasible solution for few-shot remote sensing image scene classification and demonstrates strong potential for practical remote sensing applications in which labeled training samples are scarce.Keywords：remote sensing scene classification;few-shot learning;vision-language model;hybrid metric learning3|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Soil moisture is a key regulatory factor in the energy exchange between multiple layers of the Earth system. Active and passive microwave signals provide complementary information due to their different responses to soil moisture, and integrating active and passive microwave observations to generate high-accuracy soil moisture products with excellent spatiotemporal coverage remains a current research focus and challenge. Mathematical error measurement approaches have shown potential in this integration process, as they can derive product error metrics relative to true values under certain assumptions. However, limited research has compared the performance and discrepancies of different error measurement methods when combined with different fusion algorithms, particularly in terms of fusion product accuracy. This study aims to address this research gap.Three mathematical error measurement methods, namely, the Extended Triple Collocation (ETC) method, Double Instrumental Variable (IVd) method, and Three-Cornered Hat (TCH) method, were employed to evaluate global soil moisture products derived from passive microwave-based SMAP and active microwave-based ASCAT observations, and to characterize their spatial distribution accuracy patterns. Subsequently, three main fusion algorithms (the minimum random error variance method, the maximum correlation coefficient method, and the maximum signal-to-noise ratio method) were used to generate soil moisture fusion products. The impact of different mathematical error measurement methods on the performance of fusion algorithms was then systematically analyzed.Results show the following(1) The spatiotemporal coverage of all fusion products is substantially enhanced. Among them, the combination of the TCH method and the minimum random error variance method outperforms the original single active and passive products across all error metrics (including unbiased root mean square error, root mean square error, and bias), except for the correlation coefficient (R). (2) Most fusion products exhibit superior performance compared to the ASCAT product and comparable accuracy to the SMAP product. (3) With the exception of the maximum signal-to-noise ratio method based on IVd, the fusion methods based on TCH are generally superior to those based on ETC and IVd.This study examines the impact of different mathematical error metrics on the performance of soil moisture fusion products, establishing a theoretical basis and methodological support for the fusion of active and passive microwave soil moisture products.Keywords：soil moisture;fusion algorithm;assessment method;microwave remote sensing;global scale;the minimum random error variance method;the maximum correlation coefficient method;the maximum signal-to-noise ratio method441|289|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Vegetation radiative transfer models describe the interaction between vegetation and electromagnetic waves, providing a theoretical basis for vegetation monitoring using remote sensing observations. This interaction is significantly influenced by topographic effects and leaf dorsiventrality. Accurately characterizing these two key influencing factors is of great theoretical importance for analyzing the physical mechanisms underlying the vegetation-electromagnetic wave interaction. This paper first reviews current studies on the modeling of leaf dorsiventrality and topographic effects, then summarizes our research in this field, and finally discusses the limitations of current modeling schemes and future research directions. In radiative transfer modeling, our studies have focused on leaf dorsiventrality and topographic effects, exploring their effects on vegetation absorption, scattering, thermal emission, and chlorophyll fluorescence emission, and developing leaf- and/or canopy-level radiative transfer models accordingly. This paper highlights the modeling ideas for these two influencing factors. Results show the following: (1) At the leaf scale, the asymmetry of the internal physical structure and biochemical parameters of leaves is a crucial factor that should be considered in the modeling of leaf dorsiventrality; (2) at the canopy scale, the differences in reflectance, transmittance, and illuminated fluorescence emission between the two sides of leaves need to be incorporated into dorsiventrality modeling, while plant gravitropism is an indispensable factor that cannot be ignored in topographic effect modeling. Based on the isobilateral leaf radiative transfer models and canopy models with the assumption of isobilateral leaves and flat surfaces, six radiative transfer models are developed, which are: (1) the PROLIB (PROspect + LIBerty) radiative transfer model accounting for dorsiventrality of leaf reflectance and transmittance, which working at 0.4—5.7 μm; (2) the LFD (leaf chlorophyll fluorescence with dorsiventrality) radiative transfer model that is capable of simulating dorsiventrality of leaf reflectance, transmittance, and fluorescence emission; (3) the unified optical-thermal canopy radiative transfer model 4SAID (scattering by arbitrary inclined dorsiventral leaves) accounting for the differences in reflectance/transmittance between adaxial and abaxial leaf surfaces; (4) the CFD (canopy chlorophyll fluorescence with leaf dorsiventrality) canopy radiative transfer model accounting for the differences in reflectance/transmittance between adaxial and abaxial leaf surfaces, and the differences in emitted fluorescence when the adaxial and abaxial leaf surfaces are illuminated; (5) the unified optical-thermal canopy radiative transfer model 4SAILT (scattering by arbitrary inclined leaves over terrains) considering the influences of topography and plant gravitropism; and (6) the SIFT (solar-induced chlorophyll fluorescence over terrains) canopy radiative transfer model for simulating topographic and geotropic effects on chlorophyll fluorescence. We also analyze the effects of leaf dorsiventrality and topography on canopy radiance, reflectance, emissivity, brightness temperature, and chlorophyll fluorescence based on the developed radiative transfer models and measured leaf spectra datasets. Results demonstrate that leaf dorsiventrality and topographic effects must be accounted for in vegetation radiative transfer modeling, considering their significant influences on canopy-level observations. Topography and leaf dorsiventrality simultaneously influence the interaction between vegetation and electromagnetic waves. Therefore, they need to be considered in mountainous radiative transfer modeling, which is also a key aspect of future studies. Another key research aspect is combining geometric-optical modeling theory to better represent heterogeneous canopies.Keywords：Remote sensing of Vegetation;radiative transfer modeling;leaf dorsiventrality;topographic effects;plant gravitropism21|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19Spatial scale expansion method for fractional vegetation cover based on multi-resolution tree fusion
Abstract：Fractional Vegetation Cover (FVC) is a fundamental parameter for characterizing land surface vegetation status, and achieving consistent FVC data across multiple spatial scales is crucial for comprehensive ecosystem monitoring. The Multi-Resolution Tree (MRT) algorithm is an effective method for spatial scale conversion of FVC data. However, the traditional MRT algorithm has several inherent limitations, including an insufficient physical basis for error modeling, unreasonable assumptions regarding the mean of state variables, and inadequate reliability in intermediate-scale expansion. Aiming to address these challenges, this study proposes an FVC spatial scale consistency extension method based on the improved MRT framework, integrating “scale effect error constraints, trend surface preservation, and collaborative estimation of intermediate-scale state variables.” The proposed method aims to achieve spatially consistency in scale expansion and error estimation of FVC data of varying spatial resolutions.First, this study proposes a multisource FVC scale-effect error modeling approach and introduces the resulting error into the fusion framework as the state transition error covariance. This method improves the weights of FVC fusion across different spatial scales, enhances the stability and physical consistency of spatial scale expansion, and address systematic errors caused by ignoring spatial scale differences in traditional MRT approaches. Second, to overcome the limitations of the “zero-mean” assumption for state variables that removes only random noise in traditional MRT, this study directly inputs the original FVC datasets at each spatial scale, thereby improving the realism of state estimation. Finally, a simulation mechanism for FVC at different spatial scales is established. Through weighted interpolation of high- and low-resolution FVC data and their corresponding scale-effect errors, an intermediate-scale FVC and its error estimate are jointly generated and then introduced to the MRT fusion method to achieve spatial scale expansion of FVC.The proposed method is validated in two distinct experimental regions including the Hengshui agricultural area and the Greater Khingan forest region, using Sentinel-2 and MODIS data. Experimental results show that the proposed method substantially improves the consistency between FVC data across different spatial resolutions, reducing the error to near zero in the agricultural region and from -0.056 to -0.045 in the forest region. Additionally, the proposed method can preserve most of the high-spatial-resolution details.Overall, this study presents an FVC spatial scale consistency extension method that considerably improves the consistency of results across spatial scale expansions. This proposed method also exhibits potential applicability to other related applications, serving as a methodological reference for achieving spatial scale consistency in the extension of other land surface parameters. Future work may focus on developing a multitemporal and spatial scale extension method for FVC data through the integration of temporal scale extension approaches with the proposed method in this study. This approach involves reasonably evaluating and propagating the introduced errors during the temporal reconstruction process, conducting synergistic spatiotemporal estimation to generate multiscale, spatiotemporally consistent FVC data, and assessing their estimation errors. Furthermore, angular and topographic effects can be considered to further improve the consistency of spatiotemporal FVC extension.Keywords：remote sensing;fractional vegetation cover;multi-resolution tree;scale expansion7|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Hyperspectral reconstruction (HSR) from multispectral data is an ill-posed inverse problem. Current HSR mainly focuses on the visible and near infrared spectral bands, while studies on thermal infrared HSR remain scarce due to the absence of hyperspectral thermal infrared remote sensing data, limiting the development of thermal infrared remote sensing. Simultaneously, previous studies often rely on existing satellite multispectral bands and emphasize HSR model design, while providing limited attention to multispectral partitioning and band selection strategies. This study aims to explore the ideal thermal infrared multispectral partitioning and band selection method with a certain HSR model, aiming to improve the hyperspectral reconstruction precision of thermal infrared.In this study, 727 in-situ hyperspectral thermal infrared emissivity samples were collected from seven land surface cover materials，including asphalt roads, marble, gray ground tiles, painted surfaces, green ground tiles, slate paths, and brick-concrete pavements. Broadband thermal infrared emissivity is calculated from hyperspectral thermal infrared emissivity using energy conservation law, after determining the broadband ranges using the proposed method. Several broadband partitioning methods are introduced, including the traditional equal-wavelength interval method, a cluster-based analysis method using the correlation coefficient between temperature and emissivity, and an optimization method based on Quantum Genetic Algorithm (QGA) combined with HSR model. Simultaneously, seven HSR models are employed for comparison and analysis, including non-regular multiple linear regression and stepwise linear regression (SLR), regularized ridge regression (RR), LASSO regression, and elastic network regression (ENR), nonlinear support vector machine regression (SVM) and neural network regression. These models are introduced to compare and analyze the effects of different broadband partitioning methods on HSR performance.Nonlinear HSR models and non-regularized linear HSR models have higher errors compared with the linear regularized HSR models in terms of thermal infrared emissivity HSR. LASSO and ENR models are insensitive to broadband partitioning methods, while RR is highly sensitive to broadband partitioning methods. The average error of the linear regularized RR model is the smallest, while the maximum error of ENR is the lowest for the seven land cover materials. The distribution of thermal infrared broadband affects error variation of HSR results at the wavelength direction through changes in central wavelength and bandwidth. For example, the QGA-SLR broadband optimization results reduce error differences along the wavelength direction and enhance the overall performance of the HSR model.The performance of a certain HSR model can be improved through an optimal broadband selection method, which changes with depending on the HSR model. Simultaneously, ideal thermal infrared broadband partitioning not only improves the comparability of multisource thermal infrared remote sensing products but also offers technical support for the research and development of thermal infrared remote sensing sensors. The optimized combination of broadband selection methods and HSR models provides methodological support for full-band HSR.Keywords：Hyperspectral Reconstruction;Thermal Infrared Emissivity;Broadband;Quantum Genetic Algorithm;Machine Learning Method225|103|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Models and Methods
Abstract：ObjectiveLandslide identification is critical for disaster mitigation but is constrained by heavy reliance on manual annotations, poor cross-domain generalization, and static single-pass inference that lacks self-correction. Existing methods also fail to integrate multi-source geospatial data (optical imagery, DEM, slope, aspect) or emulate the iterative reasoning of expert geologists. To overcome these limitations, we propose a dynamic, iterative, and self-refining framework that transforms static segmentation into a closed-loop reasoning process, aiming to substantially improve accuracy and geological plausibility under data-scarce and complex terrain conditions.Method We develop the Iterative-Reflective Intelligent Segmentation (IRIS) framework driven by a Multimodal Large Language Model (MLLM). It consists of four modules: (1) a multi-source data fusion layer that aligns optical, DEM, slope, and aspect features via dedicated encoders and cross-modal attention; (2) a geological reasoning module built on LLaVA-1.5-7B with geo-knowledge attention for interpreting professional descriptions; (3) an adaptive segmentation module based on SAM2-Large that produces masks under spatial prompts; and (4) an iterative reflection module that encodes previous masks as visual feedback, enabling the reasoning module to reassess outputs, check geological consistency, and refine prompts over up to T=3 iterations. The entire process is formulated as a Markov Decision Process and optimized via PPO with a multi-component reward function that combines IoU gain, geological rationality, and boundary quality, effectively alleviating the need for dense pixel-level labels.ResultOn a self-constructed Loess Plateau dataset (158 samples, 0.3—18.6 ha, three landslide types), our method achieves mIoU of 0.52, gIoU of 0.54, and cIoU of 0.53, outperforming the best baseline (Mask2Former) by 10 points in mIoU and reaching Precision@0.5 of 0.71. Compared to MLLM baselines (LISA-7B, PixelLM-7B, GeoChat), relative mIoU improvements are 85.7%, 205.9%, and 44.4%, respectively, with comparable parameters (~7.4B). Ablation studies confirm that the full reward combination yields the best performance (0.52 mIoU) and fastest convergence. Attention maps (T1→T4) visually demonstrate progressive focus refinement and boundary correction. Inference time is 2.1 s per sample, and the efficiency score (0.72) is the highest among all competitors, indicating a favorable trade-off between accuracy and computational cost.Conclusion This study demonstrates that embedding an iterative-reflective mechanism within an MLLM, coupled with multi-source data fusion and reinforcement learning, effectively overcomes the static-reasoning bottleneck in landslide identification. The significant performance gains validate that our approach emulates the expert cognitive cycle of “observation-hypothesis-verification-revision” while reducing annotation dependence. However, failure analysis shows that DEM absence (46% of failures) and inaccurate geological descriptions (31%) remain primary practical constraints, highlighting future directions: developing fault-tolerant reasoning and automated description generation. Our framework offers a new paradigm for MLLM-based geological hazard recognition, with potential transferability to debris flow and subsidence monitoring.Keywords：multi-source data fusion;remote sensing image analysis;semantic reasoning;deep learning;digital elevation model6|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Areca Yellow Leaf Disease (AYLD) is a major threat to the tropical areca industry, and its early and accurate detection is essential for effective disease management. This study aimed to evaluate the potential of multispectral remote sensing data for identifying different stages of AYLD and to analyze the sensitivity of key spectral indices to AYLD severity classification.Using the three-dimensional radiative transfer model LESS and field measurements, 65,536 simulated areca plantation plots with varying yellowing severity were constructed, from which airborne RGB and multispectral images at a 10 m × 10 m scale were simulated. First, the plots were categorized into four disease stages according to the proportion of yellowed canopy area in the RGB images. Subsequently, twenty commonly used disease-related vegetation indices were calculated from the multispectral images and aggregated to the 10m spatial resolution using two approaches: background removal (canopy-only pixels) and background included (all pixels). Finally, five machine learning models were trained with the aggregated vegetation indices and stage labels derived from RGB images to classify the four stages of AYLD. The classification performances of the two aggregation strategies and different machine learning models were then compared.The results indicated that vegetation indices calculated with background removal achieved slightly higher classification accuracy than those including background, while both approaches yielded overall accuracies exceeding 0.90. The comparable performance of the two aggregation strategies demonstrated that background mixing had a limited influence on AYLD classification at the 10 m spatial scale, supporting the potential application of multispectral data for regional disease monitoring. The models performed better for advanced stages (Stage 3 and Stage 4) than for early stages (Stage 1 and Stage 2). Field validation further demonstrated the consistency between simulated and observed data, while indicating that early-stage disease classification remained relatively challenging due to the gradual transition of spectral responses between adjacent disease stages. This was mainly attributed to the continuous changes in canopy chlorophyll content and greenness during disease progression, which made the spectral boundaries between disease stages less distinct and increased uncertainty in stage discrimination. Variable importance analysis revealed that the chlorophyll-sensitive indices TCARI and CI contributed most to the classification of disease stages, followed by green-sensitive indices such as GNDVI and GI. This indicated that canopy chlorophyll content and greenness variations were primary spectral features for distinguishing disease stages.The findings demonstrated that multispectral data at a 10 m pixel scale could effectively monitor AYLD, providing methodological guidance and technical support for rapid detection and early warning of the disease at regional scales using medium to high resolution satellite sensors such as Sentinel-2. Furthermore, this study provided a feasible framework for integrating three-dimensional radiative transfer modeling, multispectral remote sensing, and machine learning approaches for regional-scale AYLD monitoring. Such a framework could facilitate the development of efficient remote sensing-based disease assessment systems and support large-scale agricultural management applications. Future studies should incorporate multi-source remote sensing data and more detailed three-dimensional canopy information to improve the robustness and transferability of disease monitoring models.Keywords：airborne multispectral imagery;RGB imagery;three-dimensional radiative transfer model;machine learning classification;areca yellow leaf disease6|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Accurate estimation of Gross Primary Production (GPP) in mountainous forests is essential for improving our understanding of the global carbon cycle and for elucidating the responses of vegetation photosynthesis and ecosystem carbon uptake to climate variability and environmental change. However, complex mountainous terrain substantially alters illumination and viewing conditions, resulting in pronounced topographic effects and distortions in optical remote sensing observations. These terrain-induced biases can propagate into vegetation indices and consequently reduce the accuracy and spatial consistency of remotely sensed GPP estimates, particularly over steep and heterogeneous forest landscapes. To address this issue, this study integrated eddy-covariance-based GPP observations from mountainous forest ecosystems with the previously developed Topographically Corrected Near-Infrared Reflectance of Vegetation (TCNIRv) to establish a high-resolution remote sensing model for fine-scale estimation of mountainous forest GPP. Based on this model, a long-term and high-spatial-resolution GPP dataset for mountainous forests across China was generated, covering the period from 2000 to 2024 at a spatial resolution of 30 m. Results demonstrated that the developed GPP product achieved satisfactory estimation accuracy, with an R2 of 0.604 and an RMSE of 1.115 g C m-2 d-1. At the national scale, neglecting topographic effects resulted in a slight overall underestimation of mountainous forest GPP, with a mean bias of -0.176 g C m-2 d-1. Nevertheless, the influence of terrain became much more pronounced at local scales. Specifically, GPP was systematically overestimated on sun-facing slopes but underestimated on shaded slopes when topographic effects were ignored, and the magnitude of these estimation biases increased progressively with slope steepness. This study deepens the understanding of terrain-induced uncertainties in remote sensing-based estimation of mountainous forest productivity and provides a scientific basis for long-term, high-resolution monitoring of forest carbon dynamics and for improving broad-scale, efficient, and fine-grained management of mountainous ecosystems.Keywords：Gross Primary Production (GPP);mountain forests;remote sensing;topographic effects;topographically corrected near-infrared reflectance of vegetation6|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：ObjectiveAccurate estimation of the aboveground carbon stock of forests is crucial for forest management and the sustainable development of forest ecosystems. However, single machine learning models often suffer from weak generalization, low estimation accuracy, and high uncertainty when applied to forest carbon stock estimation. Aiming to address these limitations, this study aims to explore new hybrid models to improve the efficiency of model building and prediction accuracy, thereby enhancing the accuracy of aboveground forest carbon stock estimation.MethodsUsing Forest Resources Class II Survey data and Landsat 8 OLI imagery as data sources, a feature selection method integrating Genetic Algorithm (GA) and CatBoost (GA and CatBoost, GAC) was proposed. This method combines the global optimization capability of GA in exploring feature subsets with the prominent strengths of CatBoost in mining nonlinear feature relationships and quantitatively evaluating feature importance. Based on this framework, GAC was systematically compared with the traditional Recursive Feature Elimination (RFE) method to screen remote sensing feature variables that effectively improve model accuracy. Additionally, the Hyperopt hyperparameter optimization algorithm was adopted to iteratively search for the hyperparameter space of each machine learning regression model to obtain optimal parameter combinations. Subsequently, a stacked ensemble AST regression algorithm based on base-learner mean fusion and meta-learner-based adaptive weighting was constructed. The model selects four optimized single machine learning models—adaptive boosting (AdaBoost), CatBoost, Random Forest (RF), and light gradient boosting machine (LightGBM)—as base learners to fully maximize their unique advantages and markedly enhance the overall performance of the algorithm. Finally, remote sensing estimation models for carbon stock were established based on six single machine learning regression models (AdaBoost, CatBoost, RFR, LightGBM, Support Vector Machine, and Extreme Gradient Boosting) as well as the proposed AST ensemble model. After comprehensive comparisons of the estimation accuracies of all candidate models, the optimal one was selected to conduct high-precision inversion mapping of Pinus densata carbon stock in Shangri-La City.Results(1) RFE picked out 9 variables, and GAC picked out 7 variables, with the 7 variables selected by GAC contributing more to the accuracy of Pinus densata AGC inversion. (2) Using Hyperopt, the hyperparameters of each model were iteratively optimized. Results indicate that the optimal feature subset selected by GAC, when combined with the AST algorithm for regression fitting, achieved the best estimation accuracy, with a coefficient of determination R² = 0.885, a root mean square error RMSE = 8.321 t/hm², and a prediction accuracy P = 86.4%. (3) Based on the optimal estimation model, the aboveground carbon stock of Pinus densata in Shangri-La City in 2016 was estimated to be 7.70953 million t, with an average carbon density of 40.015 t/hm². (4) The directionality of texture features has a remarkable impact on the estimation accuracy of forest carbon storage based on the AST model, and the 45° diagonal direction is the optimal direction for carbon storage estimation under this model.ConclusionThe AST algorithm exhibits higher stability and anti-interference capability under multiple cross-validations, effectively improving the estimation accuracy of nonparametric models and reducing model uncertainty. This method offers an effective reference for the dynamic monitoring of forest resources in other high-altitude areas.Keywords：Carbon stocks;Hyperopt hyperparameter tuning;machine learning;AST;Pinus densata;GAC;remote sensing inversion;uncertainty analysis375|697|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Deep learning methods have demonstrated considerable potential for fine-scale extraction of plot-level crop distributions from remote sensing imagery. However, their performance depends heavily on large volumes of high-quality annotated samples, resulting in several practical limitations,including high annotation costs, poor timeliness, and limited cross-regional generalizability. These limitations are particularly pronounced in fragmented agricultural landscapes with complex plot boundaries. To address these bottlenecks, this study proposes a plot segmentation prompt optimization method that integrates a visual foundation model with agricultural prior knowledge for accurate plot-level rice mapping without region-specific model training and with minimal dependence on manual annotations. The proposed method converts rice phenological characteristics and spectral-index prior knowledge into dynamic prompts required by the visual foundation model through adaptive iterative learning.A plot segmentation prompt optimization method that couples the Segment Anything Model (SAM) with domain knowledge fusion is introduced. The proposed method is based on an adaptive iterative closed-loop system comprising four stages: “segmentation → screening → knowledge update → re-prompting.” First, prior knowledge derived from time-series vegetation indices, specifically the normalized difference vegetation index and land surface water index, is used to mask out non-vegetated areas, thereby reducing computational burden. SAM is then employed for initial segmentation. Subsequently, statistical analysis of high-confidence segmented plots is conducted to dynamically update the thresholds of prior knowledge (e.g., spectral characteristics during flooding and peak growth stages, and geometric area thresholds). These updated thresholds are automatically translated into refined prompt information, including plot center and boundary points, which are then fed back into the SAM model to guide fine-scale segmentation. Furthermore, the intersection-over-union (IoU) metric and the rate of area change serve as convergence criteria for automatic iteration termination, thereby improving algorithmic stability and computational efficiency.The proposed SAM-knowledge fusion method was systematically validated across three geographically and agronomically diverse rice cultivation regions: Ninghe District (Tianjin, China), Fujin City (Heilongjiang, China), and Niigata City (Japan). The validation results showed high mapping accuracy and cross-regional generalizability. In Ninghe, the method achieved an Overall Accuracy (OA) of 94.44%, a Kappa coefficient of 0.89, and an F1-score of 94.90%. Similarly, high performance was observed in Fujin (OA: 96.80%, Kappa: 0.91) and Niigata (OA: 94.50%, Kappa: 0.86). Comparative analysis of multitemporal imagery revealed the rice harvest period as the optimal phenological window for extraction due to maximized spectral and textural contrast. Ablation studies showed that introducing the iterative mechanism increased the Kappa coefficient from 0.79 to 0.89 compared with the baseline. Moreover, in a direct comparison with typical supervised learning models (U-Net and DeepLabV3+) trained on local samples, the proposed SAM–knowledge fusion method achieved superior extraction accuracy and boundary completeness despite the absence of training samples, demonstrating the robustness of this method under sample-scarce conditions.The proposed SAM-knowledge fusion method provides an effective solution for high-precision, plot-level rice mapping with reduced dependence on annotated samples. Through the successful coupling of the powerful, generic segmentation capability of SAM with agricultural remote sensing prior knowledge using an adaptive iterative mechanism, the proposed method overcomes the “sample dependence” bottleneck of traditional deep learning. The proposed SAM–knowledge fusion method provides a new technical solution for large-scale, low-cost, high-accuracy, and automated crop mapping, supporting food security assessment and precision agricultural management.Keywords：Remote sensing fine extraction;SAM;prior knowledge fusion;rice mapping;adaptive statistical learning;parcel-level segmentation;iterative optimization;prompt information optimization422|431|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：Particulate Organic Carbon (POC) is a key component of inland water carbon pools and plays a pivotal role in carbon transport, biogeochemical transformation, mineralization, and greenhouse gas emissions. Quantifying the concentration and source composition of POC is essential for understanding inland water carbon cycling and for enhancing regional carbon budget assessments. However, the coexistence of endogenous and terrestrial sources introduces substantial variability in optical and biochemical properties, limiting the accuracy and generalizability of available remote sensing algorithms. This study aims to apply an approach that integrates biochemical source-tracing methods with satellite remote sensing to accurately estimate POC concentration and source composition across diverse aquatic environments. A biochemical end-member mixing model was initially applied to water samples collected from the Yangtze River mainstem and Lake Taihu to quantify the relative contributions of endogenous and terrestrial POC. Based on these source apportionments, a three-band remote sensing reflectance ratio was identified as an effective optical indicator for resolving the proportion of endogenous POC. This ratio was then used to classify water bodies into two categories characterized by either dominant endogenous or terrestrial POC sources. Building upon this classification, a semi-analytical algorithm based on the inherent optical properties of the water column was applied to estimate POC concentration. Specifically, the particle absorption coefficient of pigments, aph(674), was used as a proxy for endogenous POC concentration, while the non-algal particle absorption coefficient, anap(443), was used to infer terrestrial POC concentration. The total POC concentration was then derived by combining the two source-specific estimates according to their fractional contributions. The three-band remote sensing reflectance ratio demonstrated excellent performance in estimating the proportion of endogenous POC in the water column, yielding an RMSE of 0.081, an MB of 0.0131, and a MAPE of 43.479%. The endogenous POC concentration estimated using aph(674) achieved an RMSE of 1.000 mg/L, an MB of -0.157 mg/L, and a MAPE of 24.455%. Similarly, the terrestrial POC concentration predicted that using anap(443) resulted in an RMSE of 0.346 mg/L, an MB of -0.012 mg/L, and a MAPE of 22.200%. When combined, the overall POC algorithm outperformed existing empirical and semi-analytical models, revealing an RMSE of 1.322 mg/L, an MB of -0.177 mg/L, and a MAPE of 29.380%. These results confirm that the integration of optical classification with source-specific modeling notably improves the robustness and generalizability of POC retrievals across heterogeneous inland waters. This study also demonstrates that coupling remote sensing techniques with biochemical source-tracing offers a robust framework for quantifying POC concentration and source composition at broad spatial and temporal scales. The proposed approach effectively leverages the mechanistic insights from biochemical methods and the large-scale observational capability of satellite remote sensing. The resulting algorithm enhances the interpretability and transferability of POC retrievals, offering a promising tool for advancements in carbon pool monitoring in inland waters. However, the algorithm has not yet been validated beyond the Yangtze River mainstem and Lake Taihu regions. Future work will focus on the acquisition of additional datasets from diverse hydrological and optical environment, aiming to further refine model performance and assess its extensive applicability.Keywords：inland waters;particulate organic carbon;isotopic tracing;n-alkanes;semi-analytical model350|254|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Remote Sensing Applications
Abstract：Methane (CH4) is the second most important anthropogenic greenhouse gas and contributes substantially to global warming because of its high radiative forcing and global warming potential. Effective monitoring of methane emissions is therefore essential for climate change mitigation and greenhouse gas management. Satellite remote sensing has become an important approach for methane emission detection owing to its wide spatial coverage and frequent revisit capability. Among existing satellite platforms, Sentinel-2 provides unique advantages for detecting medium- and large-scale methane point-source emissions because of its high spatial resolution and the presence of shortwave infrared bands sensitive to methane absorption. However, current methane detection studies still face several challenges, including the scarcity of real methane plume datasets, limited robustness of existing detection models under complex background conditions, and insufficient validation of model transferability across different regions, periods, and sensors.To address these limitations, this study develops a methane emission detection framework based on Sentinel-2 observations and deep learning techniques. First, a real methane plume dataset was constructed using Sentinel-2 imagery acquired over seven oil and gas emission sites located in Turkmenistan and Algeria. The dataset contains 415 satellite scenes covering diverse spatial and temporal conditions. Methane plume masks were generated through pixel-level manual interpretation assisted by spectral enhancement methods, and data augmentation was applied to increase sample diversity and model robustness. A total of 1,245 training samples were produced for model development. The dataset provides realistic methane plume characteristics and complex background information that are often absent from simulated datasets.Based on the constructed dataset, an enhanced Attention U-Net model was proposed for methane plume detection. The model incorporates spectral channel attention through squeeze-and-excitation modules to strengthen the representation of methane-sensitive spectral information and improve discrimination between methane plumes and background features. Optimized attention gates were further introduced to enhance feature transmission and suppress irrelevant signals. To address the severe class imbalance caused by the small spatial extent of methane plumes, a composite loss function combining Focal Tversky Loss, Dice Loss, and binary cross-entropy Loss was adopted, thereby improving the detection of small targets and complex plume boundaries.Comprehensive experiments were conducted to evaluate model performance from multiple perspectives. Cross-regional transfer experiments demonstrated strong generalization capability, achieving a scene-level intersection over union (IoU) of 0.917 and an accuracy of 0.926 when trained and tested in geographically distinct regions. Temporal detection experiments showed stable monitoring performance over different observation periods, with an IoU of 0.929 and an accuracy of 0.960. To assess sensor transferability, the model trained using Sentinel-2 data was applied to Landsat-8 imagery. The results achieved an IoU of 0.773 and an accuracy of 0.889, indicating promising cross-sensor application potential. Compared with the conventional MBSP threshold segmentation method, the proposed approach improved IoU by approximately 30% on average and exhibited superior robustness, noise suppression capability, and plume integrity preservation.The proposed model was further applied to real methane leakage events in Mexico and Russia. The results demonstrate that methane plumes can be successfully identified under different environmental conditions and emission intensities, including relatively weak methane signals. Overall, this study provides a high-quality Sentinel-2 methane plume dataset and a robust deep learning framework for automated methane emission detection. The proposed methodology offers valuable support for large-scale methane monitoring, greenhouse gas emission supervision, and future development of quantitative emission estimation and near-real-time monitoring systems, contributing to global climate governance and carbon neutrality efforts.Keywords：methane emission detection;Sentinel-2;methane dataset;deep learning7|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Abstract：The spatial information of cropland field parcel is an important basis for supporting agricultural resource management, production planning and scientific decision-making. The middle Yangtze River plain is characterized by interlaced water networks, diverse cropping structures, and complex surface features. Existing mainstream cropland datasets for this region mostly focus on the representation of cropland extent, while insufficiently depicting field parcel boundary details, making it difficult to meet the needs of fine-scale extraction and management.In this study, we used 1 m resolution remote sensing imagery as the main data source to construct LGISer-CFD, a cropland field parcel dataset for the Middle Yangtze River Plain. This dataset covers the Jianghan Plain, Dongting Lake region, and Poyang Lake region, and consists of two parts: a remote sensing interpretation product of cropland field parcels (LGISer-CID) and a segmentation sample dataset (LGISer-CSD). The main time phase of the image data is 2020, and the available images of the adjacent years from 2018 to 2022 are used to supplement some areas with missing images or poor quality. For LGISer-CID, the CTHBNet model was first employed to extract cropland field parcels, followed by model fine-tuning using samples from representative areas, thereby generating a cropland field parcel interpretation product covering the Jianghan Plain, Dongting Lake Plain, and Poyang Lake Plain. For LGISer-CSD, 10,000 pairs of 256×256-pixel images and labels were selected from each of the three regions, resulting in a total of 30,000 image-label pairs. These samples were divided into training, validation, and test sets at a ratio of 6∶2∶2. LGISer-CFD has the following characteristics: (1) the remote sensing interpretation product is currently the largest and highest-resolution cropland field parcel dataset for the Middle Yangtze River Plain, containing more than 5.05 million cropland field parcels and providing a comprehensive representation of the spatial distribution pattern of regional cropland field parcels; (2) the segmentation sample dataset contains 30,000 image-label pairs with a size of 256×256 pixels, which are divided into training, validation, and test sets at a ratio of 6∶2∶2. It also supports the continuous expansion of newly added field parcel samples, showing high scalability; (3) LGISer-CFD covers multiple types of cropland field parcel scenarios, including regular contiguous parcels, fragmented parcels, and parcels with curved boundaries, which meets the requirements of practical field parcel extraction tasks and can be used to evaluate model stability and generalization ability in complex scenarios. Comparisons with ten existing cropland datasets demonstrate the clear advantages of the interpretation data product in cropland extent representation, boundary delineation, and object-level segmentation. Further tests and analyses of seven deep learning models on the sample dataset show that the dataset can effectively support the training, evaluation, and comparison of cropland field parcel segmentation models, providing an important reference for future related studies.LGISer-CFD establishes an integrated cropland field parcel data system consisting of a large-scale interpretation product and a standardized segmentation sample dataset for the Middle Yangtze River Plain. By providing detailed parcel boundaries, diverse field scenarios, and standardized samples, it offers a reliable data foundation for parcel-level cropland mapping, model evaluation, and fine-scale monitoring.Keywords：cropland field parcel;high resolution remote sensing image;dataset;deep learning;the middle Yangtze River plain5|0|0<HTML><L-PDF> <Enhanced-PDF> <Meta-XML>Updated：2026-08-19
Data Articles
