最新刊期
-
摘要：Objective Fine-grained ship detection in optical remote sensing imagery is an important research topic in maritime monitoring and ocean observation. Compared with coarse-grained ship detection, fine-grained ship detection aims to distinguish specific ship types with highly similar visual appearances, which places higher demands on dataset scale, category diversity, and annotation accuracy. However, existing ship detection datasets are generally limited in the number of fine-grained categories, instance scale, and scene diversity, which restricts the performance and generalization ability of deep learning–based oriented object detection models. The objective of this study is to construct a large-scale fine-grained ship detection dataset and to further expand data diversity through controllable synthetic data generation, thereby providing a reliable data foundation and benchmark for fine-grained ship detection in optical remote sensing imagery.Method First, a large-scale fine-grained ship detection dataset, named LAFI, is constructed using high-resolution optical remote sensing images collected from 36 representative ports worldwide. The dataset contains 8,000 images acquired under diverse imaging conditions and complex maritime environments. A total of 49 fine-grained ship categories are defined, and 48,717 ship instances are manually annotated using oriented bounding boxes to accurately describe ship orientation and geometric characteristics. Second, to alleviate the limitations of real data in terms of scale and scene coverage, a controllable diffusion-based data generation framework is designed to extend LAFI into a million-scale synthetic dataset, referred to as LAFI-Diffusion. By incorporating structured textual prompts that describe scene types, weather conditions, and temporal information, the diffusion model is guided to generate realistic ship images under diverse maritime scenarios. The generated synthetic samples are further filtered and combined with real data to form large-scale training sets suitable for fine-grained ship detection. Finally, several representative oriented object detection methods are selected and evaluated on the constructed datasets to analyze the effectiveness of synthetic data augmentation and to establish benchmark results.Result Experimental results show that the proposed LAFI dataset provides improved category richness, instance scale, and scene diversity compared with existing fine-grained ship detection datasets. Moreover, incorporating synthetic data from LAFI-Diffusion into the training process consistently improves detection performance and generalization ability across different oriented object detection models. Performance gains are particularly evident in complex maritime environments, such as crowded ports and scenes with varying sea states and illumination conditions. The benchmark evaluations also indicate that the contribution of synthetic data varies across detection methods, suggesting that appropriate integration strategies are important for fully exploiting synthetic data.Conclusion This study presents a large-scale fine-grained ship detection dataset that integrates real optical remote sensing imagery with controllable diffusion-based synthetic data generation. By substantially expanding data scale and enhancing scene diversity, the proposed LAFI-Diffusion dataset effectively addresses the limitations of existing fine-grained ship detection benchmarks. The experimental results confirm that synthetic data can serve as an effective complement to real-world samples, improving detection accuracy and robustness for oriented ship detection models. The released datasets and benchmark results provide valuable support for future research on fine-grained ship detection and related remote sensing applications.关键词：remote sensing imagery;ship detection;dataset;diffusion model;synthetic data generation;Fine-grained recognition;oriented bounding box;data augmentation35|65|0更新时间：2026-04-21 -
摘要：Objective Shorelines of rivers and lakes constitute the critical interface between inland water bodies and adjacent terrestrial systems, functioning as key transitional zones that regulate flood storage, buffer hydrological extremes, and maintain aquatic–terrestrial ecological system. Accurate and timely monitoring of shoreline dynamics is therefore substantial for practical flood risk mitigation, water resource management, and the planning, protection, and restoration of freshwater ecosystems. This study aims to develop an effective and reliable deep learning (DL) model to identify shorelines from high-resolution remote sensing imagery, supporting systematic quantification of their spatiotemporal variation in response to climate and human intervention.Method Deep learning models based on the Swin-UNet architecture were set-up to identify water extents, boundaries of which were then vectorized to obtain the shorelines of the study area by using the Gaofen and Sentinel-2 remote sensing imagery respectively. The Swin-UNet was adopted in this study due to its performance in capturing both local details and long-range contextual information, enabling robust delineation of complex shoreline morphologies. Annual and monthly shorelines were retrieved from Gaofen and Sentinel-2 respectively for the period 2015–2025 based on the trained and well validated Swin-UNet models. Intra-annual and interannual variations of shorelines in terms of their length, development index, fractal dimension and composition were then assessed.Result The Swin-UNet models, based on both Gaofen (GF) and Sentinel-2 imagery, achieved a classification accuracy of water pixel exceeding 90%, with clear, continuous shoreline boundaries and well-preserved geometric details, significantly outperforming traditional thresholding and machine learning methods. Results based on Gaofen imagery show that the annual mean area of water extent in the study area was around 314.7 km², with a corresponding shoreline length of about 7,306.4 km. The results derived from Gaofen imagery were largely consistent (R2=0.99) with those from Sentinel-2. However, owing to its higher spatial resolution, the water area extracted from Gaofen imagery was about 18.9% larger than that from Sentinel-2, while the shoreline length was approximately 75% longer. Length of the shoreline showed stronger intra-annual variability (CV=12.8%) than inter-annual variability (CV=9.0%). Shorelines of artificial water bodies were found more morphologically stable than that of natural water bodies. Composition of shorelines in the study area exhibited a transition from semi-natural/semi-artificial to artificial shorelines, while natural shorelines remained relatively stable.Conclusion The study demonstrates that Swin-UNet–based deep learning approaches provide highly accurate, spatially detailed, and temporally consistent monitoring of river–lake shoreline dynamics, which is essential for effective management, protection, and restoration of inland aquatic environments. Changes in the composition of the shorelines in the study area were closely associated with urban spatial expansion and strongly influenced by a series of ecological restoration policies, including the “Grain for Green” program and wetland restoration initiatives.关键词：river and lake shorelines;dynamic monitoring;Swin-UNet;Gaofen satellite;Sentinel-2;Chenzhou City48|53|0更新时间：2026-04-21 -
摘要：Objective Deep learning methods have demonstrated considerable potential for the fine-scale extraction of plot-level crop distributions from remote sensing imagery. However, their performance is heavily reliant on large volumes of high-quality annotated samples, leading to significant challenges, including high labeling costs, poor timeliness, and limited cross-regional generalization capability. These limitations are particularly pronounced in fragmented agricultural landscapes where plot boundaries are complex. This study aims to address these bottlenecks by proposing a novel framework that achieves accurate plot-level rice mapping without the need for region-specific model training and with minimal dependence on manual annotations. By integrating visual foundation models with agricultural prior knowledge, this research seeks to establish a robust, automated solution for crop mapping in diverse environments.Method We propose a plot segmentation prompt optimization method that couples the Segment Anything Model (SAM) with domain knowledge fusion. The core of our framework is an adaptive, iterative learning closed-loop system comprising "segmentation → screening → knowledge update → re-prompting." First, prior knowledge derived from time-series vegetation indices—specifically the Normalized Difference Vegetation Index (NDVI) and Land Surface Water Index (LSWI)—is applied to mask out non-vegetated areas, substantially reducing the computational load. The SAM is then employed for initial segmentation. Crucially, the framework performs statistical analysis on high-confidence segmented plots 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 points and boundary points, which are fed back into the SAM model to guide finer segmentation. Furthermore, we innovatively introduce the Intersection over Union (IoU) metric and area change rate as convergence criteria to automatically terminate the iterations, ensuring both algorithmic stability and computational efficiency.Result The proposed 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 results demonstrated high mapping accuracy and strong generalization capabilities. 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 multi-temporal imagery identified the rice harvest period as the optimal phenological window for extraction due to maximized spectral and textural contrast. Ablation studies revealed that the iterative mechanism significantly improved the Kappa coefficient from 0.79 to 0.89 compared to the baseline. Moreover, in a direct comparison with typical supervised learning models (U-Net and DeepLabV3+) trained on local samples, our method achieved superior extraction accuracy and boundary completeness despite using no training samples whatsoever, proving its robustness against sample scarcity.Conclusion This study presents an effective framework for high-precision, plot-level rice mapping that significantly reduces dependency on annotated samples. By successfully coupling the powerful, generic segmentation capability of SAM with agricultural remote sensing prior knowledge through an adaptive iterative mechanism, the method overcomes the "sample dependence" bottleneck of traditional deep learning. It offers a promising technical pathway for large-scale, cost-effective, and automated crop mapping, providing essential data support for food security assessment and sustainable precision agricultural management.关键词：Remote sensing fine extraction;SAM;prior knowledge fusion;rice mapping;adaptive statistical learning;parcel-level segmentation;iterative optimization;prompt information optimization37|38|0更新时间：2026-04-21 -
摘要：Objective Climate change has intensified precipitation variability in arid regions, disrupting hydrological regimes and increasing flood risk. Flood control systems in these regions are frequently designed primarily for drought mitigation and water storage, limiting their capability to withstand extreme rainfall. In late July 2025, the Tumochuan Plain in Inner Mongolia experienced basin-wide flooding triggered by an unprecedented prolonged rainfall event. Multiple embankment breaches subsequently occurred along the Hasuhai drainage canal, causing severe damage to agricultural production and infrastructure. Accordingly, this study aims to provide timely and reliable information on flood evolution and quantify flood effects on croplands, particularly maize inundation and associated yield loss risk.Method A refined flood mapping method was developed by integrating a composite water index with morphological operations to enable the efficient automated extraction of surface water information. To reduce misclassification caused by terrain and building shadows, a rule-based discrimination scheme was implemented during post-processing. The proposed method was applied to Sentinel-2 imagery to extract surface water across the Tumochuan Plain for 2024 and 2025. By using water occurrence frequency in 2024, permanent and seasonal water bodies were distinguished to establish a baseline water distribution. This baseline was then used as a reference to delineate the 2025 flood extent and monitor its temporal evolution. To investigate maize growth responses to inundation duration, a dynamic time warping (DTW)–k-means model was constructed by combining DTW with the k-means algorithm to cluster the normalized difference vegetation index (NDVI) phenological trajectories. Cluster centroids represent NDVI recovery trajectories under different inundation durations. The trained model was applied across the study area to assess maize yield loss risk.Result The proposed automated water detection method outperformed the optimal threshold-based approach derived from Sentinel-1 data, achieving an overall classification accuracy of 97.4% and a kappa coefficient of 0.947. Time-series analysis indicated that flood extent peaked around August 25, with the total water area reaching 880.01 km², approximately 2.2 times the normal extent in 2024. Flood recession was slow, with the inundated area decreasing by approximately 53% over the following month. Among major crops, maize was the most severely affected, with an inundated area of 192.6 km². Approximately 39.4% of croplands remained waterlogged for more than 30 days, mostly in low-lying lands and along river systems. The DTW–k-means model clustered NDVI recovery trajectories into three types associated with flood duration. The subsequent yield loss risk assessment indicated high-risk areas that covered 238.9 km², primarily corresponding to prolonged inundation. Maize exposed to flooding for more than 2 weeks faces an elevated risk of lodging and potential mortality, indicating limited resilience to prolonged inundation.Conclusion These findings demonstrate the capability of remote sensing to link flood dynamics with abnormal crop growth responses in arid regions. The proposed framework enables rapid flood extent mapping and maize yield loss risk assessment, and it can be transferred to other regions affected by similar flood events.关键词：remote sensing;Flood disaster;automated flood mapping;crop anomaly monitoring;time-series analysis121|91|0更新时间：2026-03-27 -
摘要：(Objective)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 these two types of signals to generate high-accuracy soil moisture products with excellent spatiotemporal coverage is a current research focus and challenge. Mathematical error measurement approaches have shown great 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 these error measurement methods when combined with different fusion algorithms in terms of fusion product accuracy. This study aims to address this research gap.(Method)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 the passive microwave-based SMAP and active microwave-based ASCAT soil moisture products on a global scale, and their spatial distribution patterns of accuracy were revealed. 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 systematically analyzed.(Result)The results show that: 1) The spatiotemporal coverage of all fusion products is significantly enhanced. Among them, the combination of the TCH method and the minimum random error variance method outperforms the original single active and passive products in 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) Except for 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.(Conclusion)This study clarifies the impact of different mathematical error metrics on the performance of soil moisture fusion products, providing a theoretical basis and methodological support for the fusion of active and passive microwave soil moisture products.关键词：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 method159|104|0更新时间：2026-03-26 -
摘要：Objective The real-time and precise monitoring of crop growth status at the field scale is a critical component for achieving modern precision agriculture. Multi-source remote sensing technologies, including satellite platforms and unmanned aerial vehicles (UAVs), have emerged as effective non-destructive tools for this purpose. However, these data sources present a significant spatiotemporal trade-off, limiting their independent utility. Satellite remote sensing, while offering broad-area coverage, is often constrained by adverse weather conditions and insufficient spatial resolution to capture in-field variability. Conversely, UAV remote sensing provides exceptionally high spatial resolution but is hampered by limited battery endurance, making large-scale continuous (e.g., daily) monitoring challenging. Consequently, a single data source is inadequate for supporting the continuous monitoring of crop growth at the field scale. To address this critical gap, this study proposes a cross-platform spatiotemporal fusion method. The objective is to synergistically integrate satellite and UAV data, effectively leveraging their complementary temporal and spatial resolutions to generate a continuous, high-resolution dataset for precision agriculture.Method This research was based on a synergistic combination of multi-platform, multispectral remote sensing data, including high-resolution UAV imagery, Sentinel-2 data and PlanetScope SuperDove data. We developed an improved CACAO (Consistent Adjustment of the Climatology to Actual Observations) algorithm, adapting its core logic for cross-platform data fusion rather than its original climatological application. Two distinct data combination strategies were designed and tested: (1) a baseline “UAV+Sentinel-2” strategy and (2) an enhanced “SuperDove+Sentinel-2+UAV” strategy, which integrates high-frequency commercial satellite data. The CACAO framework was implemented using two distinct modes: a “forward prediction” (FP) mode, designed for near real-time applications, and a “backward updating” (BU) mode, which iteratively refines historical estimates as new data becomes available. The final output of the framework is a near real-time, daily 1-meter resolution normalized difference vegetation index (NDVI) time-series dataset. The accuracy of the fusion results was rigorously evaluated using two methods: (1) Leave-One-Out Cross-Validation (LOOCV), which assesses the model’s predictive power, and (2) a benchmark comparison against the established GLM-STF (Generalized Linear Model-based Spatiotemporal Fusion) algorithm.Result The prerequisite for data fusion was confirmed as NDVI data from the different platforms exhibited good consistency, with a strong correlation between Sentinel-2 and SuperDove (R = 0.97) and a reliable correlation was observed between UAV and satellite data (R > 0.75). In addition, the CACAO algorithm was proven to effectively reconstruct the phenological dynamics of the rice crop. A key finding was that the backward updating (BU) mode produced a significantly smoother and more robust NDVI time series than the forward prediction (FP) mode. Both CACAO-based data combination strategies achieved high overall accuracy (R > 0.94). Critically, the study demonstrated that introducing high-temporal-resolution SuperDove data during key phenological stages can substantially improve accuracy, with the correlation increasing from 0.51 to 0.67 in a specific validation case. Finally, in the comparative analysis, the CACAO algorithm demonstrated greater stability and slightly higher accuracy than the GLM-STF algorithm, particularly showing more robust performance across the entire growing season.Conclusion In conclusion, the cross-platform fusion framework proposed in this study, centered on the improved CACAO algorithm, is an effective and robust solution for generating continuous, high-precision (daily 1-meter) field-scale rice NDVI time series. This approach successfully overcomes the limitations of single-source data platforms. The framework provides strong technical support for the fine-grained monitoring of crop growth and the implementation of precision management strategies in modern agriculture.关键词：PlanetScope;Sentinel-2;spatiotemporal data fusion;growth monitor;precision agriculture;field scale;phenological curve;near real-time monitoring266|405|0更新时间：2026-03-17 -
摘要：Particulate Organic Carbon (POC) is a critical component of the inland water carbon pool and plays a pivotal role in carbon transport, biogeochemical transformation, mineralization, and greenhouse gas emissions. Quantifying both the concentration and the source composition of POC is essential for understanding inland water carbon cycling and improving regional carbon budget assessments. However, the coexistence of endogenous and terrestrial sources creates substantial variability in optical and biochemical properties, limiting the accuracy and generalizability of existing 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 first applied to water samples collected from the Yangtze River mainstem and Lake Taihu to quantify the contribution 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 used to classify water bodies into two categories characterized by dominant endogenous or terrestrial POC sources. Building upon this classification, a semi–analytical algorithm using inherent optical properties (IOPs) 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. Total POC concentration was then derived by integrating the two source–specific estimates according to their fractional contributions. The three–band remote sensing reflectance ratio demonstrated strong 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%. 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, terrestrial POC concentration predicted from anap(443) showed 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, exhibiting an RMSE of 1.322 mg/L, an MB of -0.177 mg/L, and a MAPE of 29.380%. These results confirm that integrating optical classification with source–specific modeling significantly improves the robustness and generalizability of POC retrievals across heterogeneous inland waters. This study demonstrates that coupling remote sensing techniques with biochemical source–tracing provides a powerful framework for quantifying both POC concentration and source composition at broad spatial and temporal scales. The approach effectively leverages the mechanistic insights offered by 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 advancing 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 acquiring additional datasets from diverse hydrological and optical environments to further refine model performance and assess its broader applicability.关键词：inland waters;particulate organic carbon;isotopic tracing;n–alkanes;semi–analytical model141|109|0更新时间：2026-03-13 -
摘要：Objective Visible light images and Synthetic Aperture Radar (SAR) images are two primary data sources in satellite remote sensing, widely used in fields such as urban planning, disaster prevention and mitigation, and national security. Feature extraction serves as a key bridge connecting remote sensing images with high-level applications, directly influencing the effectiveness of intelligent interpretation in complex scenarios. This paper systematically reviews the evolution of feature extraction methods from traditional handcrafted approaches to data-driven deep learning models, identifies the inherent challenges faced by these methods in processing visible light and SAR images, and introduces a knowledge-driven feature extraction paradigm. The aim is to explore how the integration of domain knowledge can enhance feature robustness, interpretability, and generalization, thereby advancing the field of remote sensing intelligent interpretation.Method The study follows a structured review methodology, focusing on literature from authoritative sources such as IEEE TGRS, ISPRS Journal, CVPR, and ICCV between 2019 and 2025. First, traditional feature extraction methods—including edge detection (e.g., Sobel, Canny, ROEWA), texture analysis (e.g., Gabor filters, GLCM), color features, and keypoint extraction (e.g., SIFT, Harris)—are summarized and analyzed for their strengths and limitations. Second, deep learning-based methods are examined, covering convolutional neural networks (CNNs) for local spatial modeling, graph neural networks (GNNs) for relational reasoning, and vision transformers (ViTs) for global context understanding. Third, to address challenges such as semantic ambiguity, scale variation, imaging condition sensitivity, and speckle noise interference, a knowledge-driven framework is proposed. This framework integrates three categories of knowledge—visual knowledge (e.g., edges, textures), geospatial knowledge (e.g., DEM, topological relations), and physical knowledge (e.g., scattering models, atmospheric physics)—into deep networks through mechanisms such as knowledge-as-architecture, knowledge-as-data, and knowledge-as-loss.Result The analysis demonstrates that traditional methods offer high interpretability and computational efficiency but lack adaptability to complex scenes and nonlinear relationships. Deep learning methods significantly improve feature representation and task performance but are heavily dependent on large annotated datasets and suffer from poor generalization under domain shifts. The proposed knowledge-driven approach effectively mitigates these issues. For example, incorporating edge prior knowledge enhances boundary precision in semantic segmentation; embedding elevation data improves terrain-aware classification; and integrating physical models such as electromagnetic scattering models aids in SAR image interpretation and speckle suppression. Experimental results from cited studies show performance gains across various tasks, such as increased mIoU in segmentation and higher accuracy in object detection, confirming that knowledge-guided features are more discriminative, stable, and interpretable.Conclusion Feature extraction for visible light and SAR images has evolved from manual design to data-driven learning and is now progressing toward a hybrid paradigm that combines data-driven perception with knowledge-driven reasoning. This paper highlights that integrating visual, geospatial, and physical knowledge into deep learning models can address key challenges in remote sensing feature extraction, including semantic ambiguity, scale extremes, imaging variability, and noise interference. Future research should focus on developing large-scale remote sensing foundation models, deepening the synergy between data and knowledge, and promoting cross-disciplinary integration with emerging technologies like vision-language models and diffusion models. Such advancements will further enhance the intelligence, reliability, and applicability of remote sensing image interpretation systems.关键词：remote sensing intelligent interpretation;visible light images;SAR Images;manual feature extraction;depth feature extraction;data and knowledge driven93|957|0更新时间：2026-03-13 -
摘要：Objective Crop lodging poses a significant threat to agricultural productivity and food security, yet existing monitoring approaches often suffer from low automation, insufficient integration of pre- and post-disaster data, and a lack of systematic spatiotemporal coordination. To overcome these limitations, we developed an automated framework, StandardCurve-iForest-RF, which aims to establish a resilient crop growth baseline, distinguish true lodging from noise, and enable precise spatiotemporal mapping for disaster management.Method The approach utilizes time-series Sentinel-2 satellite data to construct a Crop Growth Standard Curve (CGSC) for the target crop. The Soft Dynamic Time Warping (Soft-DTW) algorithm is employed to create this curve, which serves as a resilient reference model capable of accommodating inter-annual climatic variations. To detect lodging, the method calculates cumulative multi-feature anomaly scores by comparing post-disaster satellite observations against the pre-established standard curve. An Isolation Forest (iForest) algorithm is applied for initial anomaly detection across multiple spectral features. Subsequently, a spatiotemporal joint decision mechanism is implemented to refine the results, effectively suppressing false alarms caused by persistent cloud cover, cloud shadows, and other environmental noise. Finally, a Random Forest (RF) classifier is used to accurately map the spatial extent and precise boundaries of the lodged areas, completing the fully automated workflow from dynamic detection to precise mapping.Result The method was validated using a case study of a lodging event that occurred on September 15, 2020, in Bohetai Township, Zhaoyuan County, Daqing City, Heilongjiang Province. It successfully identified the lodging event, demonstrating its effectiveness in distinguishing actual crop damage from noise. The overall detection accuracy reached 80.36%, with a Kappa coefficient of 0.60, confirming a substantial agreement between the automated detection results and ground reference data. The results clearly showed that the integrated use of the standard curve, the anomaly scoring mechanism, and the spatiotemporal decision rules significantly enhanced the reliability of lodging identification and minimized false positives. The entire process, from data processing to the final generation of the lodging map, was executed automatically without manual intervention.Conclusion The StandardCurve-iForest-RF framework presents a significant advancement in automated crop disaster monitoring. Its core innovation lies in the construction of a resilient growth standard curve and a sophisticated spatiotemporal analysis pipeline that effectively differentiates true lodging from interference. The successful application in a real-world case study confirms the method's practical utility and accuracy. This framework provides a valuable tool for agricultural departments and emergency management agencies, enabling rapid assessment of crop damage extent and supporting timely disaster response and loss estimation. The methodology is adaptable and holds promise for application in other regions and for monitoring other types of abrupt agricultural disasters.关键词：Crop Lodging Monitoring;Standard Growth Curve;time-series analysis;Spatiotemporal Joint Detection;Sentinel-278|330|0更新时间：2026-03-13 -
摘要：Objective Accurately estimating the aboveground carbon stock of forests is crucial for forest management and the sustainable development of forest ecosystems. Since single machine learning models still face issues such as weak generalization, low estimation accuracy, and high uncertainty when estimating forest carbon stocks, 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.Methods Taking Forest Resources Class II Survey data and Landsat 8 OLI imagery as data sources, a feature selection method integrating Genetic Algorithm (GA) and CatBoost (Genetic Algorithm and CatBoost, GAC) was proposed, combining GA's global optimization capability in feature subset exploration with CatBoost's prominent strengths in mining nonlinear feature relationships and quantitatively evaluating feature importance; on this basis, GAC was systematically compared with the traditional Recursive Feature Elimination (RFE) method to screen remote sensing feature variables that effectively improve model accuracy, while the Hyperopt hyperparameter optimization algorithm was adopted to iteratively search the hyperparameter space of each machine learning regression model to obtain the optimal parameter combinations. Subsequently, a stacked ensemble AST regression algorithm based on base learner mean fusion and meta-learner-based adaptive weighting was constructed, which selects four optimized single machine learning models—Adaptive Boosting (AdaBoost), CatBoost, Random Forest Regression (RFR), and Light Gradient Boosting Machine (LightGBM)—as base learners to fully leverage the unique advantages of each model and significantly 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 (SVM), and Extreme Gradient Boosting (XGBoost)) as well as the AST ensemble model, and after comprehensively comparing 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 selected 9 variables, while GAC selected 7 variables, among which the 7 variables selected by GAC contributed more to the accuracy of Pinus densata AGC inversion; 2) Based on Hyperopt, the hyperparameters of each model were iteratively optimized, and it was found 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 significant impact on the estimation accuracy of forest carbon storage by the AST model, and the 45° diagonal direction is the optimal direction for carbon storage estimation under this model.Conclusion The AST algorithm exhibits higher stability and anti-interference ability under multiple cross-validations, which effectively improves the estimation accuracy of nonparametric models and thereby reduces model uncertainty. This method can provide an effective reference for the dynamic monitoring of forest resources in other high-altitude areas.关键词：Carbon stocks;Hyperopt hyperparameter tuning;machine learning;AST;Pinus densata;GAC;remote sensing inversion;uncertainty analysis84|309|0更新时间：2026-03-13 -
摘要：Objective Typhoon Yagi (No. 202411) is formed east of the Philippines and moved northwestward. After entering the South China Sea, it rapidly intensified into a super typhoon and brought severe winds, heavy rainfall, and flooding to the Pearl River Delta, including Guangzhou and Haikou. This study uses multi-temporal synthetic aperture radar (SAR) observations to analyze the fine-scale sea surface winds during Yagi’s evolution.Method The 500 m resolution wind products were processed with land masking, noise filtering, scalloping correction, and typhoon center detection.Result Results show that during the rapid intensification phase, both maximum wind speed and wind radii increased notably, and the winds became more asymmetric. After Yagi reached its peak intensity, these parameters stabilized. Spectral analysis based on two-dimensional Fourier transforms reveals clear kilometer-scale roll vortices in the typhoon boundary layer. Their orientations align with the outer shear flow, and the dominant wavelength ranges from 2–3 km. The spatial distribution of these rolls shows little dependence on typhoon intensity, consistent with the idea that roll formation is mainly driven by local shear instability.Conclusion Although the dataset is limited, the results demonstrate that SAR observations are effective for capturing fine-scale wind structures and boundary-layer processes. This work provides a useful reference for future large-sample studies on air–sea interaction mechanisms in different typhoon stages.关键词：synthetic aperture radar (SAR);Typhoon Yagi;Wind structure;Wind radius;Asymmetry;Boundary-layer rolls83|291|0更新时间：2026-03-13 -
摘要：Tomographic Synthetic Aperture Radar (TomoSAR) exploits multi-baseline SAR observations to synthesize an elevation aperture, enabling the retrieval of the vertical distribution of radar scatterers within a single range–azimuth resolution cell. By extending conventional two-dimensional SAR imaging into the elevation dimension, TomoSAR provides a unique capability for three-dimensional reconstruction of complex scenes, particularly in forested, urban, and cryospheric environments where volumetric or multilayer scattering dominates. Over the past two decades, TomoSAR has evolved from a primarily theoretical concept into a mature research field supported by airborne and spaceborne SAR missions, accompanied by rapid advances in inversion algorithms and application-driven studies. This paper presents a comprehensive review of TomoSAR imaging theory, height inversion methodologies, and representative application scenarios. Starting from the fundamental signal model, the TomoSAR observation process is described as an elevation-domain spectral estimation problem, where discrete and often non-uniform baselines sample the vertical reflectivity function. Based on this formulation, existing inversion approaches are systematically categorized into several major classes, including Fourier-based spectral analysis and classical beamforming, adaptive non-parametric methods such as Capon and singular value decomposition, parametric subspace-based techniques, sparse reconstruction and compressive sensing approaches, and more recent data-driven deep learning methods. For each category, the underlying principles, algorithmic assumptions, and typical implementation characteristics are summarized to provide a unified methodological perspective. A critical comparison of these methods is conducted with respect to vertical resolution, robustness to baseline sparsity and temporal decorrelation, radiometric fidelity, and computational efficiency. Conventional Fourier-based and beamforming techniques are computationally efficient and physically interpretable but inherently limited by the effective elevation aperture. Adaptive and parametric methods offer improved resolution and sidelobe suppression under favorable conditions, yet their performance strongly depends on accurate covariance estimation, model order selection, and coherence preservation. Sparse reconstruction techniques demonstrate strong super-resolution capabilities under limited and irregular baseline configurations, at the cost of increased computational complexity and sensitivity to regularization parameters. Deep learning–based TomoSAR inversion has emerged as a promising alternative, providing fast inference and adaptability to non-ideal acquisition geometries, although its generalization ability, physical interpretability, and dependence on representative training data remain active research concerns. The review further discusses key challenges that continue to limit the practical deployment of TomoSAR systems. These include baseline non-uniformity and insufficient sampling, mixed scattering mechanisms within resolution cells, temporal decorrelation and phase disturbances caused by atmospheric or platform-related effects, as well as scalability issues associated with large-area or high-resolution processing. Existing mitigation strategies, such as optimized baseline design, physics-aware modeling, phase error compensation, and regularized inversion, are summarized and critically assessed in terms of their effectiveness and limitations. Finally, emerging trends and future research directions are outlined. These include multi-frequency and multi-polarization TomoSAR for enhanced physical interpretation, hybrid inversion frameworks combining physical models with learning-based techniques, standardized validation protocols supported by external reference data such as LiDAR, and computational acceleration strategies for operational-scale applications. By consolidating theoretical foundations, methodological developments, and practical insights, this review aims to provide a coherent reference for researchers and practitioners, and to support the continued advancement of TomoSAR toward reliable and large-scale three-dimensional Earth observation.关键词：TomoSAR;tomographic height inversion algorithm;spectral estimation algorithm;compressive sensing;three-dimensional reconstruction;baseline distribution101|91|0更新时间：2026-03-13 -
摘要：Leaf Area Index (LAI) is a key biophysical parameter characterizing the canopy structure and growth status of crops. Accurate and timely monitoring of LAI using remote sensing technology is crucial for field water and fertilizer management, food security assurance, and the assessment of agricultural production potential. The red-edge region, as a spectrally sensitive band indicating leaf physiology and canopy structural changes, has been introduced on multiple medium- to high-resolution (10-30 m) satellite sensors and widely applied in crop parameter estimation, providing opportunities to further improve the accuracy of LAI retrieval. However, existing studies show considerable differences in the application of red-edge bands for LAI inversion, and the effective way to leverage red-edge information for improving LAI retrieval remains unclear due to variations in study regions. To address this issue, this study proposed a hybrid method for crop LAI retrieval by integrating the PROSAIL model with machine learning algorithms. The study developed a hybrid LAI retrieval framework by coupling the PROSAIL model with machine learning algorithms, using Sentinel-2 multispectral imagery and in situ LAI measurements of major cereal crops (rice, wheat, and maize) provided by CNERN. First, the PROSAIL model was used to generate LAI-canopy reflectance simulation dataset, and sensitivity analysis of the model parameters was conducted to clarify their contributions to the reflectance of different bands. Then, five machine learning (ML) algorithms, including Multilayer Perceptron Regression (MLPR), Support Vector Regression (SVR), Gaussian Process Regression (GPR), Random Forest Regression (RFR), and Gradient Boosting Decision Tree (GBDT), were used to develop hybrid inversion models based on the generated simulation dataset, and the optimal ML algorithm was selected based on the validation dataset. Subsequently, the performance of retrieval models using eight band combinations was evaluated to identify the optimal combination incorporating red-edge bands. Finally, based on the optimization of ML algorithm and band combination, we further evaluated the accuracy and applicability of the proposed method under different scenarios. Results demonstrated that the integration of red-edge bands effectively improved LAI retrieval performance, with the joint utilization of red-edge 1 (RE1) and red-edge 3 (RE3) contributing most significantly. MLPR performed the best in both the validation and testing datasets, highlighting its ability to capture the nonlinear relationship between canopy reflectance and LAI. Compared with the Z1 combination (Green+Red+NIR+SWIR1+SWIR2) without red-edge bands, the Z7 combination (Green+Red+NIR+SWIR1+SWIR2+RE1+RE3) achieved the highest inversion accuracy (R2 = 0.784, RMSE = 0.826), with R2 increasing 4.9% and RMSE reducing 15.6%. Further analysis indicated that the appropriate incorporation of the red-edge bands not only reduced systematic bias in LAI estimation but also effectively mitigated saturation effect in the medium to high LAI range (4 < LAI < 5), with |Bias| and RMSE decreasing by 52.2% and 41.4%, respectively. In addition, different crops showed varied responses to red-edge information, with the most significant improvement in maize (R2 increased by 17.9%, and RMSE decreased by 29.1%), followed by wheat and rice. These differences may be attributed to variations in canopy structure and leaf distribution among crop types. Overall, the combination of optimal ML algorithm and red-edge bands can significantly improve the accuracy and robustness of crop LAI inversion. This study provides methodological support and practical guidance for fully utilizing red-edge information in large-scale and long-term precise monitoring of crop growth.关键词：crop leaf area index;red-edge bands;prosail model;machine learning;band selection125|125|0更新时间：2026-03-13 -
摘要：(1)Objective Forest aboveground biomass (AGB) is a crucial component of the global carbon cycle, playing a key role in monitoring global carbon dynamics and effectively mitigating climate change. The segmentation of individual tree components to effectively distinguish the wood and leaf structures is an important foundation for accurately estimating key structural parameters of trees and precisely inferring aboveground biomass. With advancements in technology, the advent of terrestrial LiDAR has provided a novel non-destructive method for estimating tree structural parameters and AGB. However, current algorithms for individual tree component segmentation using point clouds exhibit limited universality across different tree species, and their capability to segment fine branches remains relatively constrained. (2)Method Therefore, this paper constructs a large individual tree component segmentation dataset, ITS-3D, containing 713 tree samples to address the issue of insufficient high-quality training samples for individual tree component segmentation. Additionally, the state-of-the-art Point Transformer-V3 deep learning network is employed on the ITS-3D dataset for segmentation of individual tree components including the main trunk, branches, and leaf categories. Furthermore, this paper optimizes the segmentation performance of the Point Transformer-V3 network for individual tree point clouds by constructing individual tree prior geometric features (SoD Index, PCA2 Index, and Verticality Index). Subsequently, a comparative study will be conducted with several current main deep learning algorithms. Finally, this paper also validates the effectiveness of the introduced prior geometric features through ablation experiments. (3)Result Experimental results demonstrate that the segmentation performance achieved an OA of 0.946, mAcc of 0.872, and mIoU of 0.806, reflecting high segmentation precision. Even smaller, higher-order branches within the tree crown are successfully extracted. Moreover, the model’s OA and mIoU values for various TLS subsets of ITS-3D exceed 0.928 and 0.737, respectively, showing high generalization capability across tree species with large geometric structural differences. Furthermore, in comparison with other mainstream deep learning networks, the Point Transformer-V3 network adopted in this study attains the highest values for segmentation accuracy metrics. In addition, in performance comparison with traditional leaf-wood separation algorithms such as LeWoS, the Point Transformer-V3 network increases the OA and mIoU by 7.9% and 17.2%, respectively. This fully demonstrates the excellent performance and generalization ability of high-performance deep learning techniques in individual tree component segmentation across multiple tree species. Finally, in the ablation experiments, when all prior geometric features of trees are incorporated as input, the OA and mIoU of the segmentation results reach their peak values of 94.63% and 80.59%, respectively. This represents an increase of 0.29% and 0.42% compared with the scenario where only the 3D coordinates of point clouds are used as input, indicating that introducing effective individual tree prior features during training can better promote the segmentation effect of the deep learning network. (4)Conclusion Through this research, the application of the state-of-the-art deep learning technology in individual tree component segmentation can be further improved, providing technical support for the precise estimation of tree structural parameters.关键词：Individual tree component;LiDAR point clouds;semantic segmentation;Prior features;deep learning;Leaf-wood separation;Branch extraction;Serialized attention mechanism42|97|0更新时间：2026-03-13 -
摘要：Objective Accurate estimation of leaf chlorophyll content (LCC) is of great significance for crop physiological monitoring and precision agricultural management. However, traditional vegetation indices (VIs) based on visible–near infrared canopy reflectance face notable challenges in LCC retrieval: (1) the spectral response is highly coupled with both target information (LCC) and structural noise due to the confounding effects of canopy structural signals; (2) canopy structural heterogeneity across different crop types further exacerbates the sensitivity of VIs to structural parameters, significantly limiting the generalization and applicability of inversion models for major food crops such as rice, wheat, and maize under diverse growth conditions.Method To address these issues, this study proposes a hybrid LCC retrieval framework that synergistically integrates remote sensing physical mechanisms with deep learning techniques, aiming to effectively mitigate the confounding influence of canopy structure and enhance model generalizability across diverse crop types. Specifically, a set of vegetation index ratio features (VIRFS) exhibiting low sensitivity to leaf area index (LAI) is constructed by simulating a wide range of LAI–LCC combinations using the PROSAIL radiative transfer model. Furthermore, an active learning–based strategy is incorporated into the transfer learning pipeline to optimize the selection of informative samples, thereby enabling efficient model fine-tuning with limited labeled field data. The proposed method is systematically validated on multi-crop, multi-regional datasets collected from three major agricultural zones in China: Northeast China (maize, rice, soybean), the Huang-Huai-Hai Plain (wheat), and the Yangtze River Basin (rice).Result The results show that: (1) the proposed hybrid method, integrating radiative transfer mechanisms with deep learning, achieves excellent performance in cross-regional LCC retrieval for major staple crops. It exhibits strong stability and generalizability across diverse crop types with R² consistently exceeding 0.69 and RMSE below than 4.77 μg/cm²; (2) compared to the conventional vegetation index feature set (VIFS), the newly constructed VIRFS designed to be less sensitive to LAI, significantly mitigates canopy structural interference under optimal fine-tuning conditions. Across the three major agricultural regions (Northeast China, the Huang-Huai-Hai Plain, and the Yangtze River Basin), VIRFS improves R² by 0.18–0.23 and reduces RMSE by 1.85–2.51 μg/cm² for various staple crops; (3) the transfer learning framework incorporating active learning enables high-accuracy LCC estimation using only 30% of locally labeled samples, achieving R² values of 0.69–0.74 and RMSE of 4.98–5.76 μg/cm² across different staple crops. This represents a notable improvement over random sampling–based transfer learning, with R² gains of 0.02–0.16 and RMSE reductions of 0.05–1.42 μg/cm², thereby substantially enhancing model adaptability and inversion efficiency under limited-label conditions.Conclusion In conclusion, the proposed inversion framework, which couple physical principles with data-driven methods, significantly improves the accuracy and robustness of multi-crop LCC estimation. It provides a universal solution for non-destructive LCC monitoring across diverse crops and regions, with great potential for applications in agricultural management and crop nutritional diagnosis.关键词：leaf chlorophyll content;UAV-based multispectral remote sensing;canopy structural heterogeneity;vegetation index ratio features;transfer learning;active learning73|263|0更新时间：2026-03-13 -
摘要：This study proposes a novel method for mapping landfast ice in polar regions using Sentinel-1 SAR imagery. Landfast ice, which is attached to coastal boundaries, plays a crucial role in influencing polar climate with its heat exchange and radiation processes. Due to climate change and increasing commercial activities in polar coastal areas, there is a need to map landfast ice extent using regularly acquired remote sensing SAR imagery, such as Sentinel-1. However, the scattering properties of landfast ice and surrounding rough sea ice or open water can be similar, making it challenging to distinguish between fixed and non-fixed ice regions based solely on SAR backscatter characteristics. The repeat pass InSAR method identifies according to the coherence map, which are affected by temporal decorrelation due to displacement and/or its surface melting and refreezing, particularly in early winter and/or the melting season. To address this challenge, this study introduces a method based on multiple aperture coherence (MAC), effectively improving the accuracy and temporal resolution of landfast ice identification based on Sentinel-1 imagery.The proposed method utilizes sub-aperture decomposition of single Sentinel-1 IW SLC images to generate the MAC index images. By splitting the SAR image into forward and backward sub-apertures, the coherence maps sensitive to the stability of the ice features are produced. These MAC index images are then classified into seawater and sea ice using Otsu’s thresholding algorithm. The sea ice regions are further segmented into connected components to identify landfast ice, which is defined as ice connected to the coast. The method was applied to four distinct study areas in Greenland (Melville Bugt and Jokel Bay) and Antarctica (Amundsen Coast and Banzare Coast), covering various polar environments. For validation, Sentinel-2 and Landsat imagery acquired within 24 hours of the corresponding Sentinel-1 scenes were used to assess accuracy.Validating achieved an Overall Accuracy (OA) of the proposed method of 96.84%, a Kappa coefficient of 0.9338 and an F1-score of 0.9666 (averaged across the four study sites), demonstrating the method’s effectiveness in diverse geographic conditions. The results revealed extensive landfast ice coverage in each region, with significant seasonal variations driven by environmental factors. Compared to repeat-pass InSAR, the proposed method is less susceptible to temporal decorrelation, ensuring reliable landfast ice extraction even during rapid changes (e.g., the melting season). Although a clear difference in the MAC index was observed between sea ice and seawater, the method occasionally produces sharp edges at burst boundaries in TOPS acquisitions, which may affect sea ice/water separation and warrants further investigation. Moreover, future research may also focus on applying this method to assess seasonal and interannual variations in landfast ice extent and deepening our understanding of the mechanisms influencing polar sea ice stability.关键词：polar remote sensing;landfast ice;sea ice;SAR coherence;multi aperture coherence;Sentinel-1;TOPS;InSAR49|120|0更新时间：2026-03-13 -
摘要：Hydrological connectivity is generally defined as the extent of interaction between rivers, lakes, their floodplains, or other water bodies. Changes in hydrological connectivity profoundly impact the structure and function of aquatic ecosystems, the maintenance of biodiversity, and water-related human production and livelihood activities. Human interventions such as dam, levee construction, and land reclamation, have significantly altered the connectivity in many floodplain lake regions. Therefore, exploring scientific and effective methods for quantifying hydrological connectivity, analyzing its dynamic changes, and assessing its potential ecological and environmental impacts is of great significance.The Connectivity Function ( -
摘要：Objective Global aboveground forest biomass (AGB) products have become increasingly abundant in recent years, providing valuable data for assessing carbon stocks and fluxes. However, substantial spatiotemporal inconsistencies among these products have led to large uncertainties in the estimation of global carbon storage and carbon sink strength. This study aims to systematically evaluate the interannual consistency of major global AGB products and to identify their strengths and limitations for long-term biomass monitoring and carbon accounting.Method We integrated multiple sources of AGB data, including satellite-derived products (e.g., ESA CCI Biomass, NASA JPL), dynamic global vegetation model (DGVM) simulations (NBP, cVeg), and ground-based validation datasets. A multidimensional evaluation framework was designed from three perspectives: (1) spatial consistency, assessed using correlation coefficients and spatial agreement metrics among products; (2) interannual variability, analyzed through temporal correlation and trend consistency; and (3) ground validation, performed using field observations to quantify product accuracy in regions of biomass increase and decrease.Results The results show that: (1) Different remote sensing products exhibit pronounced differences in spatial consistency. Single-epoch products based on GEDI and ICESat-2 show relatively high spatial consistency ( -
摘要：In the context of global climate warming, rising temperatures have increased both the frequency and severity of drought in subtropical regions, leading to widespread vegetation mortality and posing a substantial threat to vegetation ecosystems and forest carbon sink. Accurate quantification and understanding of vegetation drought are critical for regulating vegetation mortality rates during drought events. However, precise quantification of vegetation drought remains controversial, particularly in terms of variations among different vegetation covers and within the same vegetation cover, espeially in subtropical region with complex terrain and dense vegetations. This study aims to develop a reliable approach for the quantitative monitoring of vegetation drought in subtropical region, addressing both “same vegetation cover with different drought degrees” and “same drought degree with different vegetation covers.” To achieve precise quantification, a Vegetation Drought Response (VDR) module was developed, which characterizes the spatiotemporal response of vegetation to soil moisture over time. The module utilizes spatiotemporal features constructed from the multispectral-based modified vegetation index and land surface temperature. Drought boundaries were scientifically defined using the general nature rational that decreasing soil moisture leads to vegetation withering, while increasing soil moisture promotes vegetation flourishing, allowing identification of time intervals during the Vegetation Drought Process (VDP). Within these intervals, the sensitivity of vegetation response to soil moisture was used to determine VDR characteristics at the beginning of drought, which informs the establishment of the Vegetation Drought Threshold (VDT). By applying the VDT to VDR, a Process-Cognizant Vegetation Drought Model (PCVDM) was constructed, enabling quantitative inversion of vegetation drought. The method was applied to the Hunan-Jiangxi, which locates in the middle of China subtropical region, using remote sensing techniques to retrieve spatiotemporal changes in vegetation drought from 2000 to 2023. A spatiotemporal differentiation analysis was conducted by integrating altitude and lithology conditions to investigate causal factors. The results demonstrate that the PCVDM effectively captures the spatiotemporal dynamics of vegetation drought in the Hunan-Jiangxi region, where clear spatial differentiation is observed. High-altitude areas (>800 m) exhibit increased greenness due to rising temperatures, while low-altitude areas (<200 m) experience intensified vegetation drought as a result of enhanced evapotranspiration. Moderate-altitude areas (~400 m) show mixed responses influenced by litholog-and-slope difference, with increased greenness coexisting alongside vegetation drought phenomena. These findings suggest that vegetation drought patterns are shaped by the combined effects of climatic factors and topographic or edaphic conditions, resulting in distinct responses across elevation gradients and lithological types. The presented PCVDM provides a practical tool for remote sensing-based monitoring of vegetation drought in subtropical regions, enabling spatiotemporal differentiation across elevations and lithologies. The model reveals long-term vegetation trends in the Hunan-Jiangxi region and supports strategies for drought management, ecosystem conservation and carbon-sink study under climate warming.关键词：subtropical remote sensing;vegetation drought monitoring;process-cognizant vegetation drought model;Mann-Kendall Test;Hunan-Jiangxi Region55|103|0更新时间：2026-03-12 -
摘要：In the field of computer vision, lightweight target detection in remote sensing images is an academically challenging research topic. With the continuous improvement of the accuracy of remote sensing image target detection, the complexity of the model has increased significantly accordingly. The current remote sensing image target detection methods find it difficult to achieve an effective balance between accuracy and model complexity. Unlike conventional natural images, remote sensing images have a large imaging area, limited available information, and a multitude of target types. The vast amount of background information can easily interfere with detection, and the excessive number of target categories makes it difficult to achieve high-precision detection of specific targets. Secondly, targets in remote sensing images are densely packed, arbitrarily rotated, and exhibit significant scale variations, posing a great challenge for target localization. Furthermore, as the accuracy continues to improve in remote sensing image target detection, models have become increasingly complex in pursuit of high precision. These complex models’ heavy computational load and large parameter count severely restrict practical application on resource-constrained devices, model lightweighting has become a key factor in promoting technological application. To address the aforementioned issues, this paper proposes a lightweight remote sensing image target detection method based on Spatial-Channel Reconstruction, aiming to maintain high-precision detection of circular and rectangular-like targets while reducing model complexity. Firstly, in response to the redundant information within feature maps at the spatial dimension, a Spatial Reconstruction Unit (SRU) is adopted. It separates feature maps based on the richness of spatial features into groups with abundant spatial information and groups with redundant spatial information, and then performs cross-reconstruction operations on the two groups to reduce spatial feature redundancy and enhance the spatial feature representation of remote sensing image targets. Subsequently, to address the redundancy of channel information in feature maps, a Partial Convolution-based Channel Reconstruction Unit (PCRU) is proposed. It divides feature maps into two parts along the channel dimension, one part uses partial convolution for efficient feature extraction, and the other part uses pointwise convolution to obtain hidden detailed information. The two parts are weighted and reconstructed, and then concatenated to extract features at a lower computational cost and enhance the representation of important channels. The lightweight method based on Spatial-Channel Reconstruction is applied to ACRFNet, an adaptive circular receptive field network for remote sensing image target detection. Experimental results on the NWPU VHR-10, DIOR and HRRSD datasets verify its performance. Specifically, the method can maintain or even improve detection accuracy, while significantly reducing the model’s computational load and parameter volume. Compared with other mainstream methods for remote sensing image target detection, the proposed method demonstrates obvious advantages in balancing detection accuracy and model complexity, especially for circular and rectangular-like targets that are difficult to detect. The lightweight remote sensing target detection method based on Spatial-Channel Reconstruction effectively solves the core problem of balancing detection accuracy and model complexity. By reducing feature redundancy, it overcomes the deployment limitations of complex models on resource-constrained devices, retains good adaptability to typical targets, and provides a reliable technical support for the practical application of remote sensing target detection technology.关键词：remote sensing imagery;object detection;circular target;model lightweighting;reconstruction unit75|104|0更新时间：2026-03-09
- Postal code：100190
- Tel：010-58887052 Email：nrsb@aircas.ac.cn
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备20021838号-9
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0