Latest Issue
Online FirstList of IssuesDocument Types
Estimating the non-physiological component of canopy SIF considering dynamic leaf optical properties AI Introduction
Abstract：Solar-induced chlorophyll fluorescence (SIF) is a byproduct of photosynthesis that has the potential to rapidly reflect stress conditions in vegetation and has become an important indicator for monitoring vegetation photosynthesis. In addition to photosynthetic physiological information, observed SIF also includes non-physiological components such as leaf optical properties and canopy structure introduced during canopy radiative transfer process, making it unsuitable for direct monitoring of photosynthetic status. Existing studies indirectly calculate physiological components based on estimates of non-physiological components; however, the estimation uncertainty of non-physiological components is transferred to the calculated physiological components, thereby affecting the effectiveness of photosynthesis monitoring. Near-infrared reflectance of vegetation (NIRv) is a commonly used index for estimating non-physiological components; however, its derivation assumes that leaf optical properties are constant under ideal conditions, failing to account for changes in these properties under actual conditions such as senescence or stress. This affects the accuracy of non-physiological component estimates and consequently increases the uncertainty in the inversion of photosynthetic physiological information. To address this issue, this study proposes an estimation method for non-physiological SIF components considering dynamic leaf optical properties (NPFLeaf), based on spectral invariance theory and incorporating radiative transfer models (PROSPECT-D, SCOPE) along with field measurement data. Considering that the optical properties of cotton leaves undergo significant changes under verticillium wilt stress, cotton verticillium wilt was selected as a case study for validation in this study. The results indicate that NPFLeaf provides accurate estimates of non-physiological SIF components across various stress levels. By accounting for the dynamic changes in leaf optical properties, NPFLeaf maintains excellent performance even under severe stress conditions (R2 = 0.81) and demonstrates consistent estimation results across all stress levels (R2 = 0.85). Based on these, the physiological SIF component of the canopy calculated by NPFLeaf maintains the differences between various stress levels, and demonstrates better consistency in estimating the real physiological component across multiple stress levels, thereby helping to improve the accuracy of monitoring vegetation stress status based on changes in photosynthetic physiology. Meanwhile, the NPFLeaf estimates exhibited a strong linear relationship and good stability across different canopy structures (R2=0.97, RMSE=0.017). The relative error was lower under dense canopy conditions (7.76%) than under sparse canopy conditions (14.34%). The proposed method provides critical technical support for decoupling the physiological components of the SIF signal, helping to improve the monitoring accuracy of vegetation photosynthesis under stress and promoting the application and development of SIF in the fields of vegetation disease and growth monitoring.Keywords：remote sensing;solar-induced chlorophyll fluorescence;Spectral invariant theory;Photosynthesis monitoring;Leaf optical properties;Cotton verticillium wilt13|7|0
<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-06Research progress on the application of satellite remote sensing in disaster emergency management AI Introduction
Abstract：Emergency management is an important part of the national governance system and governance capacity, and undertakes the important responsibilities of preventing and defusing major security risks and timely responding to and handling various disasters and accidents. With the characteristics of macroscopic dynamic and long-time-series observation, especially the advantage of observation without relying on on-site personnel, satellite remote sensing has become an indispensable monitoring means to support emergency management in various countries. For security risks, satellite remote sensing can quantitatively monitor key elements and further apply them to risk identification, assessment and decision-making to support disaster early warning and risk prevention. For disasters and accidents, during the emergency period, satellite remote sensing can dynamically monitor the disaster situation to support emergency rescue and relief; after the disaster situation stabilizes, it can monitor and assess the damage to provide a reference for disaster relief, recovery and reconstruction. The application capability of satellite remote sensing in emergency management is a comprehensive embodiment of satellite observation perception capability, data analysis cognition capability and decision-making action support capability, following the perception-cognition-decision theory. To fully understand the progress of satellite remote sensing applications in emergency management, this paper explains the basic connotation and composition framework of satellite remote sensing applications in emergency management based on the perception-cognition-decision theory, and clarifies the technical process of monitoring and evaluating security risks and disasters based on various satellite remote sensing resources and supporting decision-making actions. Furthermore, this paper reviews the application of satellite remote sensing in emergency management from 2015 to 2025 from the global and Chinese dimensions in three aspects: satellite resources, risk monitoring and disaster monitoring. Overall, the global remote sensing satellite resources available for emergency management have shown explosive growth, and the observation timeliness and modalities have been continuously improved, among which China's multi-modal satellite resources have developed rapidly. In terms of risk monitoring, Europe and the United States have jointly established a number of global disaster risk monitoring and forecasting systems based on their global stable and continuous observation satellite data, while China's disaster risk monitoring based on satellite remote sensing is still dominated by local research and demonstration. In terms of disaster monitoring, the monitoring of major global disasters is a key area of cooperative application of satellite remote sensing, and a number of cooperative mechanisms for collaborative observation and rapid monitoring based on satellite resources of various countries have been formed. China actively participates in various cooperative mechanisms, and at the same time, a stable satellite application system has been established for domestic disaster response, realizing an organic connection with emergency management decision-making, and providing hourly dynamic monitoring services for emergency rescue. At present, disasters occur frequently around the world, and risks are becoming increasingly severe and complex. The spatiotemporal scale and element characteristics of satellite perception are increasingly refined, and remote sensing data are growing exponentially and being deeply integrated with artificial intelligence and high-speed computing technologies. Looking forward to the future, on the basis of innovative upgrading in sensitive perception, intelligent analysis and decision support respectively, the application of satellite remote sensing in emergency management will further realize leapfrog development in the integrated real-time monitoring of remote sensing, communication and navigation, short-term and imminent early warning of complex and extreme security risks, and people-centered actual combat service chains through systematic integrated innovation for typical emergency management scenarios, and intends to form a new application paradigm.Keywords：satellite remote sensing;disasters;monitoring and early warning;emergency monitoring;emergency management;artificial intelligence;observation-cognition29|7|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-06A general framework for monitoring and diagnosing earth surface anomaly in remote sensing images that integrates multi-dimensional features and hierarchical strategies AI Introduction
Abstract：Accelerated global climate change and intensified human activities have substantially increased the frequency and diversity of abrupt land surface disturbances, including wildfires, floods, snow disasters, algal blooms, and so on. These events are typically characterized by short temporal scales, strong physical impacts, and pronounced deviations from expected land surface states. Rapid detection, accurate spatial delineation, and reliable cause diagnosis of such anomaly are therefore essential for disaster mitigation, environmental protection, and land management. However, land surface anomaly monitoring differs fundamentally from conventional land use and land cover (LULC) change analysis. While most existing remote sensing approaches focus on long-term semantic transitions or generic spectral changes, they are not specifically designed to capture sudden, high-impact disturbances or to interpret their underlying causes. Current studies generally follow either target-specific post-event mapping paradigms, which rely heavily on predefined indices and prior knowledge, or generic anomaly detection methods that emphasize spectral deviation without semantic interpretation. Both paradigms face inherent limitations in universality, interpretability, and adaptability. This study aims to synthesize a general-purpose framework for land surface anomaly detection and diagnosis that addresses these challenges in a unified manner.This study reviews and consolidates a hierarchical remote sensing framework built upon three key components: multi-dimensional features representation, staged anomaly processing, and adaptive integration of prior information. A five-dimensional feature system is constructed to comprehensively characterize land surface physical states, representing greenness, yellowness, blackness, whiteness, and redness. These features jointly capture vegetation condition, soil exposure, water effects, snow surfaces, and combustion- or heat-related signals, forming a physically interpretable and broadly applicable feature space. The framework adopts a three-stage processing workflow. The first stage performs rapid anomaly screening to suppress background variability and highlight potential abnormal regions at large spatial scales. The second stage refines anomaly boundaries using consensus-based and weakly supervised strategies, operating exclusively on the screening outputs without introducing additional prior constraints. The third stage conducts anomaly type diagnosis by analyzing characteristic change patterns in the multi-dimensional feature space through rule-based statistical reasoning. To enhance practical applicability, the framework supports multiple operational scenarios within a unified architecture, ranging from completely prior-free conditions to cases where historical imagery or land cover information is available.Experimental evaluations across representative anomaly scenarios, including fire spots, floods, snow disasters, algal blooms, burned areas, and deforestation, demonstrate the effectiveness of the proposed framework. In the rapid screening stage, the method effectively focuses on potential anomaly regions while reducing background interference. In the refinement stage, the pseudo-label-guided U-Net produces more spatially coherent and complete anomaly regions than traditional machine learning methods, threshold-based methods, classical change detection, and representative unsupervised anomaly detection models. In the diagnosis stage, the rule-based scoring strategy distinguishes typical anomaly types using their multi-dimensional feature change patterns and further helps separate spectrally similar events such as burned areas and deforestation. Overall, the results confirm the framework’s advantages in anomaly localization, refined extraction, and interpretable diagnosis.The reviewed framework provides a robust and interpretable solution for large-scale land surface anomaly detection and diagnosis. By integrating multi-dimensional physical features with a hierarchical processing strategy and adaptive prior handling, it effectively bridges the gap between anomaly discovery and causal interpretation. The results highlight its strong generalization ability, scalability, and suitability for operational remote sensing applications, offering a promising paradigm for proactive land surface anomaly monitoring.Keywords：earth surface anomaly;abruptness;remote sensing;multi-dimensional response features;U-Net36|22|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Rapid discovery and diagnosis of sudden-onset earth’s surface anomalies from single-image remote sensing interpretation AI Introduction
Abstract：Sudden-onset earth’s surface anomalies, such as floods, landslides, and earthquakes, occur at unknown times and locations and often require response within hours. In many emergencies, the first post-event Very-High-Resolution (VHR) optical image is often the earliest available observation of the affected area, whereas pre-event reference imagery may be unavailable, imaging conditions may be uncertain, and pixel-level annotations are scarce. This work targets a practical “single-image, first-look” setting and asks whether a lightweight model can rapidly screen a single image for anomalies and, only when necessary, further delineate the affected area at pixel level with minimal supervision.We propose G2P-ESA, a unified “discovery-diagnosis” framework that converts a single VHR image into a semantic-guided attributed graph and then translates graph evidence into prompts for a foundation segmentation model. First, a pretrained superpixel segmentation network generates superpixels as approximate geo-entities and provides deep appearance features, while a pretrained land-cover segmentation model supplies a semantic prior. For each superpixel node, pooled deep features are concatenated with geometric centroid coordinates and a one-hot semantic embedding to form node attributes; edges are built via k-nearest neighbors with distance pruning to obtain a compact, noise-controlled graph. Second, a lightweight Principal Neighbourhood Aggregation graph neural network performs image-level anomaly discovery. The classifier outputs the normal/anomalous probability and can be extended to anomaly-type prediction when training labels are available; the anomaly confidence further controls whether downstream diagnosis is triggered, avoiding unnecessary computation on normal scenes. Third, under weak supervision, Graph-GradCAM++ is used to estimate node importance for the predicted target class and identify key evidence nodes. These nodes are post-processed using thresholding, connectivity completion, isolated-cluster removal, and semantic consistency filtering to generate high-confidence foreground/background prompts. Finally, the prompts drive a frozen Segment Anything Model (SAM) to generate candidate masks; an iterative refinement strategy selects an intermediate-confidence mask and re-prompts SAM once to improve boundary completeness and reduce over- or under-segmentation.We construct MESA, a large-scale multi-class single-image earth’s surface anomaly benchmark with 20351 samples from 84 events, including normal scenes and 12 anomaly categories, and further build a pixel-level subset MESA-Seg with 240 images for diagnosis evaluation. Experiments show that G2P-ESA achieves 94.92% overall accuracy on the MESA test set for discovery, and reaches 65.31% IoU and 75.22% Dice on MESA-Seg for pixel-level diagnosis. On the public ESADv2 benchmark, where all anomalies are merged into a single “abnormal” class for fair comparison, the discovery branch yields strong binary screening performance while remaining lightweight. The discovery branch contains only 0.78M trainable parameters, supporting fast screening under constrained compute and storage budgets; the diagnosis stage uses a frozen SAM and is activated only for anomalous scenes.G2P-ESA provides an interpretable, low-supervision, single-image solution for emergency remote sensing that unifies rapid anomaly screening and pixel-level delineation. By representing images as semantic-guided graphs, using GNN reasoning for robust discovery, and translating graph evidence into prompts for SAM, the framework shortens the processing pipeline and reduces annotation dependence. The results indicate that graph-driven prompt generation is an effective pathway toward practical first-look screening and on-demand diagnosis in time-critical Earth observation.Keywords：emergency remote sensing;earth’s surface anomaly;single-image remote sensing;graph neural network;weakly supervised semantic segmentation45|21|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Research on AI-Empowered Technical Pathways for Immediate Remote Sensing of Sudden Earth Surface Anomalies AI Introduction
Abstract：Immediate remote sensing of sudden earth surface anomalies represents a major strategic demand for high-quality national economic and social development and has also become an emerging research frontier in geoscience, remote sensing, and related fields. Focusing on the key technical bottlenecks constraining satellite remote sensing of sudden earth surface anomalies, this study explores new concepts, mechanisms, paradigms, and technical pathways from the perspective of the convergence between artificial intelligence (AI) and advanced satellite remote sensing technologies, and proposes an overall AI-empowered solution for immediate remote sensing of sudden earth surface anomalies. First, to address the challenge of “failing to detect”, AI-enabled technical pathways are proposed for autonomous and coordinated satellite scheduling, detection of small anomalies/targets, and perception of information under occlusion. Second, to address the challenge of “failing to respond rapidly”, AI-enabled pathways are developed for restructuring satellite remote sensing task chains, innovating data transmission mechanisms, and accelerating anomaly perception and processing. Third, to address the challenge of “failing to identify accurately”, AI-enabled approaches are proposed for incorporating physical constraints and knowledge guidance, improving the recognition accuracy of weak anomaly signals, and optimizing data processing. On this basis, the major technical challenges facing future AI-empowered immediate remote sensing of sudden earth surface anomalies are analyzed from the perspectives of data acquisition, algorithms and models, on-orbit intelligent computing and communications, and overall system architecture. Finally, future development directions are discussed, including the transition from task-specific models to foundation models and AI agents, from perceptual intelligence to cognitive intelligence, and from post-event response to pre-event early warning.Keywords：sudden earth surface anomalies;immediate remote sensing;AI empowerment;on-orbit edge computing29|16|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Sensitivity analysis of hyperspectral image compression to geometric and radiometric errors for on-orbit processing AI Introduction
Abstract：Hyperspectral remote sensing imagery provides rich spectral and spatial information and plays an important role in on-orbit Earth-surface anomaly detection and rapid diagnosis. However, actual satellite imaging is often affected by geometric misregistration, radiometric disturbances, stripe noise, and other imaging errors. Existing deep-learning-based compression methods generally assume that inputs are geometrically aligned and radiometrically stable, and their robustness under realistic on-orbit imaging conditions has not been adequately investigated. This study aims to analyze the sensitivity of hyperspectral image compression to geometric and radiometric errors, clarify how these errors propagate through deep compression networks, and develop a robust training strategy for reliable on-orbit hyperspectral compression and processing.A unified observation model was established to describe geometric and radiometric errors in hyperspectral images. Geometric errors were modeled as band-to-band spatial misregistration, whereas radiometric errors were represented by additive and multiplicative disturbances. Based on this model, the propagation of imaging errors through the encoder, latent representation, quantization, entropy model, and decoder was analyzed. Four representative deep compression models—mbt2018, GMM, STF, and BTN—were evaluated under varying geometric, radiometric, and mixed-error conditions. Degradation in spatial reconstruction and spectral fidelity was quantified using PSNR- and SAM-related metrics. To improve robustness, a knowledge-distillation-based training strategy was introduced. A teacher network trained on clean images was used to guide a student network trained on error-corrupted inputs, enabling the student model to learn more stable latent representations while maintaining rate-distortion performance. Downstream scene-level and pixel-level classification experiments were also conducted to assess how compression and imaging errors affect semantic discriminability.The experimental results show that geometric and radiometric errors affect hyperspectral compression in different ways. Geometric misregistration primarily disrupts spatial-structural consistency and weakens cross-band correspondence, leading to marked degradation in reconstruction quality. Radiometric disturbances have a stronger effect on spectral fidelity and cause greater spectral distortion, particularly at high error intensities. Under mixed-error conditions, the knowledge-distillation-based strategy mitigated compression-performance degradation more effectively than clean-data training and direct mixed-noise training. In the quantitative comparison, the proposed strategy reduced ΔPSNR and ΔSAM to 1.9 dB and 1.5°, respectively, indicating greater robustness than the baseline training schemes. The downstream classification results further indicate that compression may amplify the adverse effects of imaging errors on semantic interpretation. For scene-level classification, knowledge distillation increased accuracy from 82.78% to 86.11% under error-corrupted conditions. For pixel-level classification, the overall accuracy increased from 78.31% to 87.62%, demonstrating improved preservation of fine-grained spectral discriminability.Geometric and radiometric imaging errors can be significantly amplified during deep hyperspectral image compression and may further affect downstream anomaly-detection and classification tasks. The proposed knowledge-distillation-based training strategy effectively suppresses error amplification, improves compression robustness, and preserves semantic information under realistic on-orbit imaging conditions. These results provide useful guidance for robust hyperspectral image compression, intelligent on-orbit processing, and reliable Earth-surface anomaly detection.Keywords：hyperspectral remote sensing imagery;deep learning-based compression;knowledge distillation;on-orbit processing;Earth surface anomaly detection;error propagation41|19|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Knowledge graph-guided single-image surface anomaly diagnosis AI Introduction
Abstract：This study aims to enable rapid discrimination of surface-anomaly occurrence and diagnosis of anomaly type from a single post-event remote sensing image when pre-event imagery and event-specific auxiliary data requiring real-time acquisition are unavailable during emergency response. It targets seven representative disaster types—earthquake, fire, flood, typhoon, landslide, tsunami, and volcanic eruption—and specifically addresses the recurring problem of “different events exhibiting similar spectral or textural signatures” in single-temporal observations.We construct a domain-specific geoscience knowledge graph structured around “surface anomalies-affected elements-environmental factors.” Long-term, stable, offline-available fundamental geographic datasets (e.g., administrative boundaries, climate, elevation, and population) are used to encode multiscale spatial co-occurrence relationships and directional semantics as structured priors. Building on this graph, we design a dual-branch feature-fusion network (MFD-NET) that aligns visual representations with graph-based semantic reasoning in a unified semantic space and performs adaptive gated fusion. We also introduce a neighborhood-consistency-based posterior calibration mechanism to impose semantic constraints and enhance diagnostic stability, thereby mitigating single-image ambiguities such as ‘different events with similar spectral or visual signatures.’ The knowledge graph organizes adjacency, containment, and directional relationships among anomalies, affected elements, and environmental factors. It is jointly embedded with CNN-derived visual features through linear projection and temperature scaling. Four neighborhood co-occurrence consistency metrics are defined to quantify agreement between predictions and graph-based priors. These metrics are incorporated into an adaptively weighted classification loss and a consistency-driven KL-divergence constraint, which jointly guide model training and posterior correction.Experiments on a representative natural-disaster remote sensing dataset demonstrate that the proposed method achieves competitive accuracy and stability in both tasks. It achieves an overall accuracy of 94.2% for anomaly-occurrence discrimination (binary classification). For anomaly-type diagnosis (7-class classification), producer’s and user’s accuracies remain high across categories, with limited variability under small-sample conditions. Even when the training ratio is reduced to 20%, the overall diagnostic accuracy remains above 90% with only limited variability. Knowledge-guided fusion notably reduces confusion between morphologically or environmentally similar categories, such as flood–landslide and earthquake-landslide. Compared with representative baselines based on MLPs, Transformers, and state-space models, MFD-NET achieves higher overall accuracy while maintaining competitive inference efficiency. Supplementary experiments on the Sen1Floods11 dataset yield an overall accuracy of approximately 92% for flood discrimination, indicating reasonable transferability to medium-resolution multispectral conditions.By integrating offline geoscience priors through a knowledge graph and coupling them with visual evidence through semantic alignment, gated fusion, and posterior calibration, the proposed approach provides a feasible pathway for rapid surface-anomaly diagnosis from a single image without relying on pre-event imagery or event-specific real-time auxiliary data. It also offers practical support for emergency remote sensing interpretation and intelligent image understanding. The current knowledge graph mainly covers seven typical natural-hazard categories and supports diagnosis at the image and entity levels. Future work will extend the ontology to additional anomaly types and compound events and pursue pixel-level localization to further improve spatial granularity and applicability.Keywords：Knowledge Graph;Single remote sensing image;surface anomaly type diagnosis;lightweight reasoning;feature-level fusion;emergency remote sensing14|12|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Simulation study of immediate remote sensing detection mechanisms for earth-surface anomalies using an integrated communication-navigation-remote sensing satellite constellation AI Introduction
Abstract：Immediate remote sensing represents a major advance in Earth observation by dramatically shortening the closed-loop cycle from observation through understanding to action. It shifts remote sensing from post-event analysis toward intervention during an event and even preemptive warning. Abrupt surface anomalies, characterized by rapid onset, rapid evolution, and severe consequences, are among the most direct targets for immediate in-orbit remote sensing detection. However, realizing immediate remote sensing is challenging because it depends on the deep integration and coordinated development of multiple emerging technologies. In addition to integrated “communication-navigation-remote sensing” constellation networking, these technologies include high-speed intersatellite and satellite-to-ground data links, novel sensor payloads, onboard processing and AI chips, space-based distributed artificial intelligence, and in-orbit anomaly-detection and diagnostic algorithms. A dedicated mechanism is also required to coordinate the detection workflow and integrate these technologies. Under current or moderately advanced capability conditions, this mechanism should minimize the overall workflow duration while maintaining detection effectiveness.Research on these mechanisms requires specialized simulation tools. Existing simulation software primarily reproduces conventional technologies and is inadequate for simulating emerging technologies still under development, modeling the outcomes of specific algorithms, and designing specialized process sequences. This study investigates immediate remote sensing detection mechanisms for land-surface anomalies using an integrated “communication-navigation-remote sensing” satellite constellation. It first defines the scope and constraints of the mechanism study, then examines the distinctive requirements for simulation software, and finally develops a simulation tool to conduct a comprehensive analysis of the key modes of immediate remote sensing detection. For six key mechanisms, including space-based parallel scheduling and on-demand algorithm upload, batch-based quantitative effectiveness evaluations were conducted using the EM-DAT disaster-case dataset. The results validated the timeliness of the selected detection modes and technical mechanisms and provided a feasible technical framework and software tools for further in-depth research on immediate remote sensing detection mechanisms.This study randomly selected 800 disaster cases from 2022 to 2025 as detection targets, with 200 cases from each year. Simulation results for the six key mechanisms show that, under the current satellite-constellation architecture, the proposed mechanisms can achieve anomaly detection within approximately ten minutes to half an hour, with all mechanisms substantially reducing detection time. Enhancing single-satellite detection capability, including the use of single-scene diagnostic algorithms or payloads with stronger detection capabilities, is the most direct and effective means of achieving immediate remote sensing detection because it avoids communication and transmission delays. In scenarios requiring interaction, including task scheduling, information fusion, and result transmission, space-based intersatellite interaction is more efficient than satellite-to-ground interaction.Although the experimental conditions were relatively idealized, the study demonstrates the substantial potential of research on immediate remote sensing mechanisms. The experimental results further support the validity and necessity of development directions such as space-based computing, onboard intelligence, and satellite-cluster collaboration. Future research should, first, gradually transition from simulations under ideal conditions to real-world environments and, second, integrate optimization theory and artificial intelligence into the design of mechanisms and workflows. In this setting, satellites and ground stations participating in the detection process can be modeled as AI agents capable of autonomous decision-making. This approach would enable automated simulation across diverse scenarios and initial conditions while allowing key mechanisms to emerge autonomously during the simulations.Keywords：earth’s surface anomalies;immediate remote sensing;detection mechanism;satellite constellation simulation;satellite constellation architecture;communication-navigation-remote sensing integration;in-orbit processing27|14|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05The comprehensive measurement framework and method for remote sensing early warning of sudden earth’s surface anomaly intensity AI Introduction
Abstract：In recent years, various types of sudden earth’s surface anomalies have occurred with increasing frequency worldwide, posing substantial threats to socioeconomic sustainability and ecological security. During emergency response, merely identifying the spatiotemporal occurrence of such anomalies is no longer sufficient to meet practical needs. Rapid characterization of anomaly intensity is increasingly required to support immediate early warning and emergency decision-making. Owing to its broad spatial coverage, high spatial continuity, and repeat observation capability, satellite remote sensing has become an important means for assessing earth’s surface anomaly intensity over large areas. However, existing remote sensing approaches for anomaly intensity assessment often rely on large volumes of auxiliary data and computationally complex algorithms, making it difficult to achieve the immediate generation of warning knowledge for multiple types of earth’s surface anomalies under the constrained storage and computing resources available on-orbit satellites. To address this limitation, this study develops a comprehensive measurement framework and methodology for remote sensing early warning of sudden earth’s surface anomaly intensity. Based on the multidimensional representation mechanisms of anomaly intensity, the proposed framework systematically organizes features at the pixel, structural, object, and scene scales, thereby establishing a general and hierarchical measurement system applicable to multiple types of earth’s surface anomalies. On this basis, a lightweight method integrating multi-feature information is developed for the immediate generation of remote sensing warning knowledge on earth’s surface anomaly intensity, providing methodological and technical support for on-orbit intensity assessment and immediate early warning of sudden earth’s surface anomalies.Keywords：sudden earth’s surface anomaly;intensity;immediate remote sensing;immediate early warning;comprehensive measurement framework30|17|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Earth surface anomaly detection using spectral-semantic feature fusion and graph neural networks AI Introduction
Abstract：Earth surface anomaly (ESAnom) is abnormal phenomena occurring on the Earth’s surface that are often associated with severe natural or anthropogenic disasters. Rapid and accurate detection of these anomalies is crucial for disaster response, risk mitigation, and environmental monitoring. However, existing ESAnom detection methods suffer from two major limitations. First, many approaches rely on multi-temporal or multi-source satellite data, which introduces inevitable delays and limits their applicability in time-critical scenarios. Second, most methods are designed to detect a single anomaly type, requiring multiple task-specific models to handle diverse ESAnom categories, thereby increasing computational complexity and hindering large-scale deployment. To address these challenges, this study aims to develop a unified and efficient framework capable of detecting multiple ESAnom types simultaneously from a single satellite image.This paper proposes a novel ESAnom detection framework, termed SpecSemGNN, which leverages graph neural networks (GNNs) to integrate spectral and semantic information for single-image, multi-class anomaly detection. The framework consists of three main components. First, a land-cover parsing module extracts high-level semantic features and class-aware representations from the input satellite image. Second, a dual-graph construction module fuses predefined spectral indices with semantic features and constructs two complementary graph structures: a spatial adjacency graph to model local neighborhood relationships, and a semantic correspondence graph to capture inter-class and intra-class associations. Third, a dual-graph reasoning module performs message passing and inference over the constructed graphs, enabling the model to exploit both spatial and semantic dependencies for robust ESAnom identification. To support model training and evaluation, a new large-scale dataset named Sen2ESA was constructed. This dataset covers 135 major ESAnom events worldwide from 2016 to 2025 and includes seven representative anomaly categories: building damage, wildfire, flood, oil spill, red tide, volcanic eruption, and algal bloom.Extensive experiments were conducted on the Sen2ESA dataset to evaluate the effectiveness of the proposed approach. SpecSemGNN was compared against representative CNN-based, Vision Transformer (ViT)-based, and GNN-based methods under consistent experimental settings. Quantitative results demonstrate that the proposed method consistently outperforms competing approaches across multiple evaluation metrics. In particular, SpecSemGNN achieved the highest overall detection accuracy of 89.37%, exceeding the strong GNN baseline ViG-s by 1.14%, while also delivering superior per-class F1 scores for most ESAnom categories. Qualitative analyses further indicate that the integration of spectral and semantic features significantly improves class separability and robustness, especially for anomalies with similar visual appearances or spectral characteristics.The experimental results validate the effectiveness of the proposed SpecSemGNN framework for Earth surface anomaly detection. By jointly leveraging the physical interpretability of spectral indices and the discriminative power of semantic representations within a graph-based reasoning framework, the proposed method enables accurate, efficient, and generalizable multi-class ESAnom detection from a single satellite image. This work provides a practical and scalable solution for real-time disaster monitoring and lays the foundation for future research on fine-grained anomaly localization and diagnosis.Keywords：earth surface anomaly;immediate remote sensing;graph neural network;spectral index;disaster detection23|17|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Design of a knowledge-structured system for remote-sensing-based early warning of surface anomalies AI Introduction
Abstract：With the growing impacts of climate change and intensified human activities, surface-anomaly events have become more frequent and complex, making timely responses increasingly difficult. Existing remote sensing monitoring systems are often designed for post-event analysis, and a semantic gap persists between physical observations from remote sensing and the disaster-assessment standards used by emergency management agencies. This study aims to construct a structured knowledge system for remote-sensing-based early warning of surface anomalies, thereby supporting the transformation of multisource observational data into standardized warning knowledge and automated decisions.A “data-information-knowledge-decision” framework was designed to organize the early-warning process into four progressive layers: monitoring information, anomaly information, warning information, and decision information. Based on the shared physical responses of different surface anomalies, a cross-domain mapping mechanism linking industry standards to remote sensing parameters was established. A three-dimensional quantitative assessment model was developed, comprising anomaly intensity, impact, and trend. Intensity was represented by the Land Surface Anomaly Intensity index, which integrates pixel-, structure-, object-, and scene-level remote sensing features. Impact was evaluated using a rule-based discrete-grid model that couples anomaly intensity with exposed elements such as populations and assets. Trend was predicted by combining grid dynamics with a BiLSTM-CNN model to characterize the spatiotemporal evolution of anomalies. Furthermore, EM-DAT records, satellite imagery, and supplementary textual information were integrated to construct a remote sensing early-warning knowledge graph and a dual-modal knowledge base that supports historical deduction, dynamic updating, and rule-based reasoning.The proposed system provides a unified semantic structure for heterogeneous surface-anomaly events and converts continuous remote sensing parameters into standardized warning levels ranging from 0 to 4. Case studies of forest fires in California, USA, and Yunnan, China, demonstrated the feasibility of the proposed framework. The multiscale intensity assessment produced a clear spatial gradient of anomaly severity and provided direct quantitative input for warning decisions. The grid-based impact assessment identified the spatial distribution of affected populations and high-priority response areas. The trend-prediction model captured the potential spread direction and stage of fire evolution, enabling the knowledge base to trigger warning decisions automatically. The knowledge-graph case further showed that historical disaster records and remote-sensing-derived indicators could be jointly organized and used for reasoning within the same event-centered structure.This study establishes a structured knowledge framework and a technical pathway for remote-sensing-based early warning of surface anomalies. By linking disaster-management standards with remotely sensed physical parameters, the proposed system helps overcome data fragmentation and semantic inconsistency across disaster types and management sectors. The framework provides a theoretical and technical basis for knowledge-driven, multihazard, and collaborative early warning and offers a scalable foundation for the future development of intelligent surface-anomaly monitoring and rapid on-orbit warning applications.Keywords：land surface anomaly;real-time remote sensing;disaster early warning;structured framework;Knowledge Graph;early warning knowledge base19|16|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05End-to-end simulation and quantitative evaluation of on-orbit remote sensing imagery for the “Muduo” satellites AI Introduction
Abstract：The “Muduo” remote sensing satellites are designed to support rapid surface-anomaly detection and onboard intelligent processing. Before launch, however, the lack of observations acquired by the target payloads constrains payload-performance assessment, algorithm development, and predeployment testing. This study therefore aims to establish a physically consistent, end-to-end image-simulation method for the “Muduo” 1A multispectral and 1B hyperspectral payloads. The method is intended to generate prelaunch data that reproduce the principal radiometric, spectral, and spatial characteristics of the target sensors, thereby supporting subsequent system evaluation and algorithm validation.EnMAP hyperspectral imagery, comprising 224 bands at a spatial resolution of 30 m, was fused with four 10 m Sentinel-2 bands using the Gram-Schmidt method. Because Sentinel-2 lacks a panchromatic band, an equivalent panchromatic component was constructed from its blue, green, red, and near-infrared bands. The fusion produced a surface-reflectance base map containing 224 spectral bands at a spatial resolution of 10 m. The Second Simulation of the Satellite Signal in the Solar Spectrum (6S) model was then used to convert surface reflectance to top-of-atmosphere (TOA) radiance under specified atmospheric and observation conditions. Look-up tables were introduced to improve computational efficiency. Sensor spectral responses were simulated by convolution with normalized spectral-response functions, whereas spatial degradation was represented using a Gaussian point-spread function derived from the modulation transfer function at the Nyquist frequency. Frequency-domain filtering and spatial resampling were subsequently applied to generate images matching the spatial characteristics of the “Muduo” 1A and 1B payloads. Because actual “Muduo” observations were unavailable, Landsat 8 OLI data were used for surrogate validation. Simulated Landsat 8 TOA radiance was compared with temporally adjacent official Landsat 8 observations using root-mean-square error, mean absolute percentage error, Pearson correlation coefficient, structural similarity, and peak signal-to-noise ratio. Additional module-comparison experiments examined the respective contributions of atmospheric radiative transfer and spatial-response simulation.The simulated and reference images showed generally good radiometric agreement and preserved the principal spatial patterns of farmland, forest, and forest-fire scenes. In the Hebei farmland experiment, the Pearson correlation coefficient in the near-infrared band reached 0.883. Across the three validation scenes, the maximum correlation coefficient was 0.942, structural similarity reached approximately 0.90, and peak signal-to-noise ratio exceeded 25 dB for most evaluated bands. Removing the 6S atmospheric module substantially increased radiometric errors; for example, the mean root-mean-square error for the Hebei farmland scene increased from 12.74 W/(m2·sr·μm) to 28.97 W/(m2·sr·μm), while structural similarity decreased from 0.790 to 0.616. Removing modulation-transfer-function filtering produced smaller but consistent reductions in spatial fidelity. The simulated imagery also reproduced the expected resolution differences among the 20 m “Muduo” 1A data, the 10 m visible–near-infrared data from “Muduo” 1B, and its 30 m shortwave-infrared data. The two payloads therefore exhibited complementary capabilities for representing spatial structure and characterizing diagnostic spectral information.The proposed method provides a physically based simulation chain from surface reflectance to target-sensor observations and maintains useful radiometric and spatial consistency under the tested conditions. It can support prelaunch payload analysis, onboard algorithm testing, calibration studies, and application planning for the “Muduo” satellites. Because the current assessment relies on surrogate Landsat 8 observations and a limited number of representative scenes, further validation and model refinement should be undertaken when actual “Muduo” in-orbit data become available.Keywords：EnMAP;Muduo satellite;data simulation;6S;SRF;spatial response degradation;spectral simulation;surface anomaly12|15|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05A spatiotemporal prior knowledge-driven method for real-time generation of direct impact knowledge of sudden surface anomalies AI Introduction
Abstract：Sudden land-surface anomalies such as wildfires, floods, earthquakes, volcanic eruptions, and hurricanes may rapidly alter or obscure the original land-cover state. In emergency response, however, pre-event high-resolution normal-state imagery is often unavailable in a timely manner, while a single post-event image alone cannot reliably reveal which land-cover types were affected or quantify their degrees of impact. This study aims to generate direct impact knowledge—including affected land-cover types, spatial distributions, and impact degrees—from a single post-event remote-sensing image and its associated acquisition time and geographic location, thereby reducing dependence on pre-stored pre-event imagery.A two-stage framework integrating historical spatiotemporal prior generation and direct impact knowledge generation is proposed. In the first stage, a latent diffusion model with a Diffusion Transformer backbone learns normal-state land-surface distributions from historical Sentinel-2 imagery. Temporal information is encoded using sine and cosine functions of the normalized day of year, while geographic information is represented using an improved position-encoding strategy that accounts for image size and spatial resolution. Given the location and acquisition time of a post-event image, the model generates a 10 m normal-state image. A reference-guided super-resolution model then uses the post-event 0.6 m image as structural guidance, particularly exploiting information from unchanged regions, to reconstruct a spatially aligned 0.6 m pre-event normal-state image. In the second stage, the generated normal-state image and the observed post-event image are input into a bi-temporal affected land-cover identification network to estimate pre-event land-cover semantics and anomaly-affected regions. Their combination produces the affected land-cover map, and a lightweight convolutional network further estimates the degree of impact for each land-cover category. Three datasets were constructed using xBD and crowdsourced data: a historical dataset of approximately 70000 Sentinel-2 images, a super-resolution dataset containing 17963 image pairs, and a direct impact dataset containing 20 disaster cases and 200 annotated image pairs.The spatiotemporal prior model achieved an overall MAE of 0.055 and FID of 1.186, indicating good consistency between the generated and observed normal-state images. The reference-guided super-resolution model obtained an overall SSIM of 0.523 and FID of 67.444. For affected land-cover identification, the proposed framework achieved an OA of 0.836, mIoU of 0.630, and overall F1 score of 0.745. Its performance was second only to the upper-bound setting using real pre- and post-event high-resolution images, and it outperformed approaches using generated or real low-resolution pre-event images. The method performed particularly well for buildings, reaching an F1 score of 0.827 and an impact-degree MAE of 0.105. Ablation experiments confirmed the effectiveness of joint spatiotemporal conditioning, improved geographic encoding, and post-event-image-guided super-resolution. The complete model required approximately 2.7 GB of storage, and the average processing time was about 20.02 s per square kilometre on an NVIDIA P40 GPU, indicating potential for deployment in resource-constrained environments.Generative models can compress historical spatiotemporal priors into model parameters and reconstruct a task-oriented pre-event reference when real pre-event high-resolution imagery is unavailable. The proposed framework provides an effective route from a single post-event observation to structured direct impact knowledge, supporting rapid affected-object identification and impact assessment under storage- and data-constrained remote-sensing application scenarios.Keywords：remote sensing;sudden surface anomaly;disaster response;anomaly direct impact knowledge;single-scene image;generative model;spatiotemporal encoding10|14|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Semi-supervised learning based method for building anomaly detection and evaluation in remote sensing images AI Introduction
Abstract：Frequent earth surface anomalies pose severe threats to human life and property, necessitating the development of rapid and accurate detection and assessment methods. As a core element of artificial surfaces, buildings are key targets for surface anomaly monitoring. However, constrained by high annotation costs, sparse anomalous samples, imbalanced class distributions, and noisy pseudo-labels, existing deep learning models often suffer from significant performance degradation on target domains, failing to meet the “few-shot, fast-response” application requirements. This paper proposes a novel semi-supervised learning framework that leverages a small set of labeled samples alongside abundant unlabeled data to enhance model generalization for building anomaly detection and assessment.We propose a Dual-view Teacher and Category Adaptive Semi-supervised Learning method (DTCA). Built upon the robust teacher-student consistency learning framework, DTCA introduces four key improvements tailored for building anomaly analysis: (1) Spectral-Geometric Differentiated Strong Augmentation: distinct geometric perturbations (e.g., shear, stretch, shift) are applied only to the pre-event image and its labels, while varied color augmentations are applied to both pre- and post-event images, expanding the perturbation space and improving robustness to spatial misregistration. (2) Dual-view Teacher Network: two teacher networks with asynchronous momentum updates are fed with differently weakly-augmented views of the same sample; their predictions are averaged to generate more stable and reliable pseudo-labels, enhancing knowledge integration from sparse anomalies. (3) Category-Adaptive Soft Threshold: a soft thresholding function based on per-class statistics (mean and variance of prediction confidence) is constructed per class to assign dynamic weights to pseudo-labels, adaptively suppressing noisy labels while preserving medium-confidence samples, addressing class imbalance. (4) Prediction Discrepancy-Guided Re-training: the discrepancy between the two teachers’ predictions is quantified as a per-sample uncertainty score; highly uncertain samples near decision boundaries are selected into a re-training set, effectively mitigating class imbalance and improving learning from difficult instances.Extensive experiments were conducted on three public and challenging building anomaly datasets: xBD (optical, multi-class damage assessment), BRIGHT (cross-modal optical-SAR), and TUE-CD (optical, binary change detection). Under various labeling ratios (5%, 10%, 20%), DTCA consistently and significantly outperformed six state-of-the-art semi-supervised learning methods. For instance, on xBD with only 5% labeled data, DTCA achieved an overall F1 score of 71.13%, surpassing the supervised baseline by 11.83% and even exceeding the model trained with 20% labeled data (70.15%). The advantages held across different modalities and task formats. Ablation studies confirmed the individual contribution of each proposed component, and further analyses on parameter sensitivity, pseudo-label filtering strategies, and different network architectures validated the general applicability of DTCA.This paper presents DTCA, a novel semi-supervised learning method for building anomaly detection and assessment. By integrating spectral-geometric differentiated augmentation, a dual-view teacher network, category-adaptive pseudo-label filtering, and uncertainty-guided re-training, DTCA effectively leverages unlabeled data to boost performance with limited annotations. Experiments demonstrate its superiority over existing methods and its compatibility with different modalities, task formulations, and networks. DTCA offers a practical and efficient solution for rapid model fine-tuning and generalization where labeled anomaly data is scarce. Future work may explore integrating foundation models or generative techniques for further cost reduction and pre-emptive model generalization.Keywords：remote sensing imagery;deep learning;semi-supervised learning;consistency learning;earth surface anomaly;building22|29|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Intelligent modeling of surface anomaly trends driven by grid-based dynamics simulation： theoretical methods and practical applications AI Introduction
Abstract：In recent years, abrupt surface anomaly events, such as dust storms, landslides, and marine oil spills have become frequent, requiring high-precision and near-real-time prediction for early warning. However, traditional physical models are computationally complex and strongly data-dependent, whereas pure data-driven models lack physical consistency and are difficult to deploy on satellite platforms. To address these challenges, this study proposes a lightweight intelligent modeling framework for surface anomaly trends driven by grid dynamics simulation. This framework adopts a ground-training and on-board-inference paradigm to enhance accuracy, efficiency, and cross-scenario generalization in simulating multiple types of surface anomalies.This study first constructs a universal grid dynamics model applicable to solid, liquid, and gaseous phase surface anomalies. This model simulates landslide movement, oil spill diffusion, and dust storm propagation by discretizing surface space into grid cells and designing local evolution rules based on physical driving mechanisms, thereby generating a physically consistent multimodal simulation dataset. Then, a lightweight BiLSTM-CNN spatiotemporal fusion network is designed for surface anomaly trend prediction. Specifically, a double-layer stacked BiLSTM captures bidirectional temporal dependencies, while depthwise separable convolutions with dual spatial-channel attention extract multi-scale spatial features. Additionally, a gated fusion mechanism dynamically fuses spatiotemporal information.The proposed grid dynamics driven intelligent modeling framework effectively bridges the gap between computationally intensive physical models and data-driven approaches often lacking physical consistency. The performance of both the grid dynamics model and the BiLSTM-CNN network were validated across several independent representative surface anomaly case studies. In the solid phase landslide scenario, using the 2019 Shuicheng landslide in Guizhou as a case study, the grid dynamics simulation achieved a Sørensen coefficient of 0.88, a Kappa coefficient of 0.89, and an overall accuracy of 94.12%. For the deep learning model, the test set yielded a mean squared error (MSE) of 0.001 and a structural similarity index (SSIM) of 0.96. In the liquid phase oil spill scenario, validation used DeepSpill experimental data alongside historical spill cases from the Gulf of Mexico and Dalian Xingang. The simulation achieved an overall accuracy of 96.70% with a Kappa coefficient of 0.91, while the deep learning model yielded an MSE of 0.001 and an SSIM of 0.98. In the gas phase dust storm scenario, simulations of dust events in Northwest China showed that the diffusion direction was highly consistent with observed wind-field evolution. The deep learning model achieved an MSE of 0.001 and an SSIM of 0.97 on the test set. These results indicate that the proposed framework effectively characterizes complex spatiotemporal evolutionary features of diverse surface anomaly events.The proposed grid dynamics driven intelligent modeling framework effectively bridges the gap between computationally intensive physical models and data driven methods lacking physical consistency. By integrating physical process simulation with deep learning, this approach achieves high accuracy and robust universality across diverse surface anomaly predictions. Furthermore, the lightweight architecture establishes a viable technical path for real-time on-orbit monitoring and early warning. Future research will focus on incorporating more refined physical priors and optimizing model structures through knowledge distillation, pruning, and quantization to further enhance adaptability in spaceborne computing environments.Keywords：immediate remote sensing;earth’s surface anomalies;grid dynamics;landslide;oil spill21|14|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Knowledge expression mechanisms and methods for immediate remote sensing early warning of sudden earth surface anomalies AI Introduction
Abstract：Immediate remote sensing of sudden earth surface anomalies is a frontier field that leverages the rapid response capabilities of remote sensing technology to enhance large-scale disaster prediction and early warning. A critical bottleneck in this domain is efficiently transforming anomaly data detected by satellite imagery into early warning knowledge. Overcoming this is essential for shifting away from traditional single-image transmission modes and deepening on-orbit feature extraction, knowledge discovery, and real-time semantic representation. The key to solving this problem lies in establishing a knowledge representation model and an immediate conversion mechanism from spatial data (image) to natural language (text). Grounded in the advanced concept of immediate remote sensing, this study first proposes the modal characteristics and structural system for representing earth surface anomaly early warning knowledge. Subsequently, addressing the demands for knowledge-oriented, immediate, lightweight, and intelligent on-orbit processing, this paper explores the immediate conversion pathway from spatial monitoring data to semantic warning knowledge under resource-constrained on-orbit conditions. The methodology includes: (1) constructing a semantic element framework and extraction system encompassing anomaly location, extent, intensity, impact, and trend; and (2) proposing a method for the immediate generation of early warning messages that integrates the early warning expression knowledge graph with a lightweight large language model (LLM). Furthermore, a satellite-ground collaborative control strategy is employed to ensure the stability and professionalism of the generated results. Tests on the NVIDIA Jetson AGX Orin platform further showed that the complete processing time was generally maintained within one minute under the 50 W power mode. For the computationally demanding Hurricane Milton case, the total processing times were 51.9 s under 50 W and 77.4 s under 15 W. The lightweight language model was compressed to approximately 993 MB and achieved a generation speed of 5.1tokens/s, indicating the feasibility of deploying the proposed method in resource-constrained edge environments. Ultimately, this study establishes an immediate early warning knowledge expression technology system for earth surface anomalies under ‘resource-constrained, high-efficiency, and multi-user’ conditions. This system provides crucial methodological support for the remote sensing early warning of sudden earth surface anomalies.Keywords：sudden earth surface anomalies;immediate remote sensing;early warning knowledge expression;Knowledge Graph;Large Language Model (LLM);semantic generation;on-orbit processing17|10|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Few-shot earth surface anomaly detection via large vision-language models AI Introduction
Abstract：Earth Surface Anomaly (ESAnom), including various natural and human-induced hazard events, pose significant threats to public safety and socio-economic stability. Remote sensing has become a key technology for large-scale and timely ESAnom monitoring due to its wide coverage and rich observational capabilities. However, practical ESAnom detection remains challenging because of the diversity of anomaly types, large intra-class variations, and the scarcity of high-quality labeled samples. Conventional deep learning methods usually rely on extensive annotated datasets and often suffer from severe performance degradation under few-shot settings. The objective of this study is to develop a robust and efficient few-shot learning framework for ESAnom detection by leveraging large vision-language models, enabling accurate anomaly recognition with limited labeled remote sensing data.This paper proposes a lightweight few-shot ESAnom detection framework built upon vision-language pretrained models such as CLIP and RemoteCLIP. The framework consists of three key components: text prompt generation, cross-modal few-shot adaptation, and collaborative inference. First, a unified prompt template is designed and combined with a large language model to automatically generate diverse and semantically rich textual descriptions for each ESAnom category, alleviating semantic insufficiency caused by limited image samples. These descriptions are encoded by the text encoder to form structured semantic representations. Meanwhile, labeled remote sensing images are encoded by the frozen image encoder to obtain visual features. Second, a lightweight cross-modal adapter is introduced to jointly model image and text features in a shared embedding space. The adapter contains a linear mapping layer and a classification head, whose weights are initialized using class-wise image feature prototypes to improve training stability under few-shot conditions. During training, both image and text features are treated as supervised samples, enabling effective visual-semantic alignment. Finally, a collaborative inference strategy is designed for the testing stage, where the supervised predictions from the adapter and the zero-shot predictions derived from text prompts are fused via weighted integration, combining task-specific discriminative information with the generalization capability of pretrained vision-language models.Extensive experiments are conducted on two ESAnom datasets (ESAD and DisasterM3) and two conventional remote sensing scene datasets (AID and NWPU-RESISC45). Results under 1-shot to 16-shot settings demonstrate that the proposed method consistently outperforms state-of-the-art few-shot approaches, including Linear Probe, CLIP-Adapter, and Tip-Adapter variants, across different base models. Notably, the performance gains are more pronounced on ESAnom datasets, highlighting the effectiveness of incorporating rich textual semantics in complex anomaly scenarios. Ablation studies further verify the contributions of each component, showing that the proposed adapter structure, cross-modal training strategy, and collaborative inference mechanism all play essential roles in performance improvement. Visualization results confirm that the proposed adapter significantly enhances feature compactness and inter-class separability.This study presents an effective few-shot ESAnom classification framework that integrates vision-language pretrained models with lightweight cross-modal adaptation and collaborative inference. By exploiting automatically generated textual semantics and aligning them with limited visual samples, the proposed method achieves robust and scalable performance under both few-shot and larger-sample settings. The framework not only reduces reliance on large annotated datasets but also provides a flexible solution for diverse and complex ESAnom scenarios. Future work will explore more fine-grained semantic modeling and incorporate temporal remote sensing observations to further enhance robustness in dynamic and evolving anomaly detection tasks.Keywords：earth surface anomaly;remote sensing;few-shot learning;large language-vision models;cross-modal learning18|14|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05A spectral shape encoding-based self-supervised method for burned area mapping AI Introduction
Abstract：Rapid and accurate remote sensing monitoring of wildfires is essential for disaster emergency response, burned area mapping, ecological damage assessment, and post-fire vegetation recovery analysis. Conventional burned area detection methods mainly rely on spectral indices derived from reflectance magnitudes in specific bands. Although these methods are convenient for identifying fire-related changes, they neglect the physical information embedded in the overall spectral shape and are sensitive to threshold selection, vegetation conditions, soil background, and regional environmental differences. Consequently, their accuracy and robustness are often limited when burned areas are spectrally similar to bare soil or non-photosynthetic vegetation. This study aims to develop an effective feature extraction and self-supervised classification framework for multispectral imagery to improve burned area detection under complex surface conditions and limited labeled samples.To overcome the strong dependence on spectral amplitude and the difficulty in capturing holistic fire-induced spectral perturbations, this study proposes a self-supervised burned area detection method based on spectral shape encoding. Sentinel-2 multispectral imagery is used as the primary remote sensing data source, and six representative wildfire study areas distributed across North America, South America, Africa, and Asia are selected to evaluate the applicability of the proposed method under different geographical and ecological conditions. Spectral samples obtained from public spectral libraries and interpreted satellite imagery are combined to construct the experimental dataset. A total of 3,114 spectral samples are collected, including 1,316 samples from the ECOSTRESS, FRAMES, and United States Geological Survey spectral libraries and 1,798 samples extracted from the study areas. The proposed approach integrates Ordinal Pattern Encoding （OPE）, Asymmetry Encoding （AE）, and Spectral Angle Encoding （SAE） into a hybrid spectral encoding strategy that transforms Sentinel-2 multispectral reflectance into a 23-dimensional morphological feature vector. OPE characterizes local fluctuations and ordinal relationships among adjacent spectral bands, AE measures local asymmetry and smoothness, and SAE quantifies gradient relationships between representative bands. By combining these complementary descriptors, the method explicitly converts spectral shape information into physically interpretable features. A SimCLR-based contrastive learning framework is subsequently employed for self-supervised pretraining. Different augmented views of the same spectral sample are used to learn discriminative feature representations by increasing the similarity between positive sample pairs while distinguishing different samples in the feature space. The pretrained encoder is then applied to a downstream four-class classification task using a limited number of labeled samples.The proposed method achieves a Dice Coefficient （DC） of 88.34%, a Commission Error （CE） of 14.15%, and an Omission Error （OE） of 9.02%, demonstrating better overall performance than conventional spectral indices and several machine learning and deep learning baselines. Compared with classification using the original spectral reflectance, which achieves a DC of 81.36%, the proposed spectral encoding framework improves the DC by 6.98 percentage points. It also outperforms representative comparison methods, including random forest, support vector machine, CatBoost, ResNet50, VGG-16, and U-Net. Ablation experiments further reveal significant complementarity among OPE, AE, and SAE. The complete hybrid encoding achieves a DC of 88.34%, representing an improvement of 29.43 percentage points over the best single-encoding configuration, OPE, which achieves a DC of 58.91%. The experimental results indicate that the integration of multiple spectral shape descriptors effectively improves the discrimination of burned areas from spectrally similar surface types and contributes to stable classification across different study regions.These results demonstrate that explicit spectral shape modeling and contrastive learning can improve the accuracy, robustness, and physical interpretability of burned area detection from multispectral imagery. The proposed framework reduces reliance on manually selected spectral thresholds and extensive labeled datasets while providing an effective strategy for feature extraction and self-supervised classification. This study offers a practical spectral morphology-driven approach for wildfire remote sensing monitoring and provides methodological support for burned area identification and complex land surface analysis.Keywords：spectral shape;spectral encoding;burned area;self-supervised classification;remote sensing monitoring;contrastive learning;multispectral imagery;feature extraction11|17|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05A relationship-mapping-driven method for instantaneous knowledge generation on multi-dimensional indirect impacts of earth surface anomalies AI Introduction
Abstract：Remote sensing technology can efficiently identify direct physical damages caused by earth surface anomalies (such as wildfires, heavy rain, and earthquakes), but providing immediate, quantitative, and interpretable assessments of the secondary and cascading impacts is still a core challenge in the field of emergency management.This paper proposes an instant generation framework for indirect impact knowledge that integrates multi-source remote sensing proxy variables and heterogeneous relationship mapping. The framework adheres to the principles of “object-driven, model-adaptive.” First, it systematically decouples the differentiated impact pathways of disasters on four types of vulnerable entities: ecology, population, farmland, and transportation, and constructs clear mechanistic dedicated proxy variable systems for each. Next, it designs and builds the most suitable heterogeneous relationship mapping models for the response mechanisms of various vulnerable entities, including an ecological environment impact knowledge generation model based on indicators and states, a population impact knowledge generation model based on exposure-vulnerability relationships, a crop impact knowledge generation model based on process-output relationships, and a transportation impact knowledge generation model based on Bayesian inference. Rather than adopting a one-size-fits-all modeling strategy, the framework flexibly employs linear aggregation, statistical fitting, nonlinear learning, and probabilistic graphical reasoning according to the characteristics and complexity of each impact pathway. Ultimately, by assimilating real-time remote sensing data into the corresponding models, the framework generates spatial distribution maps of risk levels, statistical loss reports, and other quantitative knowledge products.To validate the effectiveness and universality of the framework, this study selects three typical earth surface anomaly events: the California wildfires, the “7.20” heavy rain in Henan, and the earthquakes in Puerto Rico for empirical analysis. These events are representative of different hazard types, providing a robust testbed for the framework's generalizability. The case results show that: (1) From an ecological perspective, the framework quantifies the impact levels of ecosystems and their spatial differentiation after wildfires; (2) At the population level, it accurately identifies high-risk areas and scales of affected populations in earthquake events; (3) In agriculture, it successfully estimates crop yield loss rates based on single-phase data; (4) In traffic, it assesses the probability of traffic network connectivity failure and the degree of impact caused by earthquakes. These results collectively indicate that this framework not only achieves high-precision evaluations but also systematically reveals the underlying chain patterns of “earth surface anomalies-processes-impacts,” which are difficult to capture directly through traditional remote sensing.This framework can transform direct damage information into indirect impact knowledge with clear physical significance, bridging the gap from remote sensing observational data to the assessment and generation of indirect impact knowledge for earth surface anomaly events. By enabling rapid and evidence-based assessments within the critical emergency response window, it provides a new methodological pathway for understanding the cascading impacts of earth surface anomalies and supports more timely emergency response, risk management, and resilient disaster risk governance.Keywords：remote sensing;Earth surface anomalies;indirect impacts;proxy variables;causal inference;knowledge generation11|10|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-05Remote Sensing Emergency Monitoring of the Venezuela Earthquake AI Introduction
Abstract：Objective The consecutive strong earthquakes that struck northern Venezuela in June 2026 caused widespread coseismic surface deformation, building damage, traffic disruption, and secondary geological hazards. This study aims to develop a multi-source remote sensing emergency monitoring approach that integrates rapid regional-scale detection with detailed identification of key affected targets, thereby providing spatial information support for post-earthquake rescue deployment, restoration of transportation lifelines, and prevention of secondary hazards.Method Sentinel-1 SAR imagery, Sentinel-2 multispectral imagery, and high-resolution optical images from Vantor/Maxar, Beijing-3, and Jilin-1 were jointly used to establish a multi-source collaborative monitoring workflow consisting of regional deformation detection, disaster extent identification, and detailed interpretation of key targets. Coseismic line-of-sight deformation was retrieved using differential interferometric synthetic aperture radar (DInSAR). Severely damaged buildings were extracted using the bi-temporal Building Damage Grading Inspection Network (BDGINet). Road blockages were identified through a road-constrained bi-temporal change detection method combined with road-segment determination. Post-earthquake landslides were extracted using the Landslide Extraction Network (LENet), which integrates pre- and post-event Sentinel-2 imagery with slope information. The deformation, building damage, road blockage, and landslide results were further integrated within a unified spatial framework for statistical and spatial association analyses.Result Clear coseismic deformation signals were detected from all four Sentinel-1 tracks. The deformation zone extended for approximately 200 km along the fault system in northern Venezuela, and the maximum line-of-sight displacement near the fault exceeded 0.3 m. A total of 261 severely damaged buildings were identified in the coastal urban belt of La Guaira. Forty road blockage sites were detected across four administrative regions, with a cumulative blocked length of 2190.60 m. The precision, recall, and F1-score of road blockage identification reached 87.50%, 92.11%, and 89.74%, respectively. In addition, 31, 75, 52, and 63 landslides were identified within the VIII–IX, VII–VIII, VI–VII intensity zones and outside the intensity VI zone, respectively. The monitoring results indicate that the spatial distribution of earthquake damage and secondary hazards was jointly controlled by fault rupture, seismic intensity, topographic and site conditions, infrastructure exposure, and structural vulnerability.Conclusion By integrating coseismic deformation, building damage, road blockage, and post-earthquake landslides into a unified multi-source remote sensing emergency monitoring framework, this study establishes a cross-scale assessment chain progressing from regional deformation detection to detailed identification of key exposed targets. The proposed framework improves the completeness of earthquake disaster-chain information acquisition and enhances the spatial linkage among different disaster products. It can provide a rapidly implementable technical approach for delineating priority search-and-rescue areas, restoring transportation lifelines, screening landslide risks, and supporting post-earthquake recovery and reconstruction. However, the monitoring results are still affected by cloud cover, image spatial resolution, cross-sensor differences, and the regional transferability of the models. Further validation and updating are therefore required using higher-resolution remote sensing imagery, UAV surveys, and field investigations.Keywords：Venezuela earthquake;multi-source remote sensing;DInSAR;building damage;road blockage;landslide;emergency monitoring;disaster analysis83|27|0<HTML><L-PDF><Enhanced-PDF><Meta-XML>Updated：2026-11-02
