ABSTRACT Employing active learning is a viable approach to reducing the human effort required to complete geographic entity recognition tasks using deep learning methods. Nevertheless, current sampling strategies often only capture one sample characteristic (e.g. information content) and disregard the temporal dynamics of the sample feature across different iterations. Moreover, active learning methods alone can only alleviate workload at the sample selection level. To address these issues, we propose a novel active learning framework that integrates temporal features into sampling and leverages a large language model (LLM) to enhance annotation efficiency and the precision of model prediction. Specifically, we propose a mixed sampling strategy based on temporal and semantic similarity, comprising three innovative indicators: temporal instability, dynamic variance entropy, and semantic similarity. We propose an approach to utilizing the LLM Qwen2.5-7B-Instruct for machine annotation and an expert system for annotation correction, further reducing the workload of geographic entity recognition tasks. Experiments demonstrate that our proposed sampling strategies outperform random sampling, with “temporal instability” exhibiting optimal performance and robustness. Furthermore, combining machine annotation with expert systems can save approximately 27% of annotation time. Our work provides a practical solution for designing sampling strategies with temporal features and fully leveraging LLM for automatic annotation to reduce the annotation cost of geographic entity recognition tasks.

ABSTRACT Semantic change detection (SCD) aims to identify land cover changes between bi-temporal remote sensing images and plays a crucial role in applications, such as urban monitoring and disaster assessment. Most existing methods rely on a shared-weight encoder with a dedicated change extractor, which limits their ability to precisely localize changed regions and maintain intra-class semantic consistency. Although vision foundation models (VFMs) exhibit strong generalization ability, their potential for SCD remains largely underexplored. In this paper, we introduce ChangeVFM, a novel yet effective framework that unleashes the power of vision foundation models for semantic change detection in remote sensing images. This framework is supplemented by a spatiotemporal modeling module to capture fine-grained spatial details and a query-based feature injector that integrates VFM’s semantic priors with multi-scale spatiotemporal features. The feature injector ensures that the ChangeVFM excels in both maintaining semantic consistency and multi-scale information of bi-temporal images. Without bells and whistles, ChangeVFM achieves competitive performance on the HRSCD, SECOND, and Landsat-SCD. Comprehensive quantitative and qualitative experiments further validate the effectiveness of the introduced modules and the robustness of the proposed method.

ABSTRACT Terrain-neglected radiometric distortion remains a major challenge in quantitative remote sensing over slope terrain. While numerous topographic correction models have been developed and validated in specific regions or simulated settings, a systematic and quantitative evaluation of the satellite-derived reflectance bias introduced by terrain neglect is still lacking. A global whole-year multi-angular reflectance data set was produced for >15000 homogeneous 500 m single-slopes by coupling MODIS daily products with a physically based radiative-transfer model, including both terrain-neglected and terrain-corrected surface reflectance. The relative percent bias of terrain-neglected reflectance to terrain-corrected reflectance (rBias) was evaluated across four slope groups (10°–15°, 15°–20°, 20°–30°, ≥30°). The results show that rBias is primarily governed by slope-driven variation in relative solar zenith angles (rSZA). Positive and negative values of rBias flips with rSZA: terrain-neglected reflectance is totally overestimated at small rSZA (~0°–15°) while totally underestimated at high rSZA (~69°–90°), and the whole proportion of underestimation increases with slope, from 58.4% at 10°–15° to 71.5% at ≥30°. Magnitude of rBias escalates non-linearly with both slope and rSZA, as median rBias ranges from −29.5% to 4.2% at 10°–15° slopes and from −61.2% to 11.9% at slopes ≥30°. Also, with the rBias extrema expand from (−55.2%, 24.8%) at 10°–15° slopes to (−88.3%, 55.0%) at slopes ≥30°. These findings quantify terrain-neglected satellite-derived reflectance bias and highlight the need for topographic correction under steep slopes and high rSZA.

ABSTRACT Fractional vegetation cover (FVC) is a critical biophysical parameter that quantifies the proportion of green vegetation projected vertically onto a unit ground area. It serves as a fundamental indicator for monitoring ecosystem health, modeling land surface processes, and assessing environmental changes such as desertification and soil erosion. While satellite remote sensing has become the dominant method for large-scale, high spatial resolution FVC monitoring, significant challenges persist in complex mountainous regions due to topographic effects and heterogeneous vegetation patterns, which complicate the validation of FVC products in these areas. This study derived high spatio-temporal resolution reference true-value products (RTVPs) for FVC in a typical mountain area through the synergistic integration of in-situ measurements, unmanned aerial vehicle (UAV) observations, and the Sentinel-2 constellation. The approach involved establishing a multi-temporal dataset of high-resolution UAV-based FVC true value data through space-air-ground synchronous observation experiments, developing a terrain-aware random forest regression model incorporating multi-dimensional features including surface reflectance, vegetation indices, topographic factors, observation geometry, and image texture, and constructing a spatio-temporal continuous FVC dataset through the harmonic modeling of Sentinel-2 like 10 m datasets. Validation showed that our UAV-scale FVC retrieval achieved an R 2 of 0.9623 and an RMSE of 0.0508 using the pixel dichotomy method. The mountain-specific FVC retrieval model demonstrated exceptional performance with an R 2 of 0.9406 and an RMSE of 0.0598 with the UAV reference maps. The resulting FVC RTVPs provide 10 m spatial resolution with a 5-d temporal resolution, effectively capturing fine-scale vegetation dynamics while maintaining temporal continuity. These RTVPs offer unprecedented accuracy for validating existing fine and coarse spatial resolution FVC products and serve as a benchmark for ecological modeling in complex terrain.

ABSTRACT Graph convolutional networks (GCN) excel in rapidly mining semantic information from data with superior capabilities, enabling deep feature extraction and classification for hyperspectral images (HSI). However, supervised classification methods heavily rely on many high-quality labeled samples, which is often a significant limitation in real-world applications. Unsupervised learning, with its unique ability to automatically discover potential structures in data without labeled samples, provides a new approach to address this issue. Thus, this paper proposes an unsupervised dual-masked graph autoencoder feature extraction network (UDMG) for HSI classification. This method first utilizes a masked autoencoder to mask and reconstruct the edges and node features of the graph structure. Meanwhile, the encoder combines GCN and Transformer to capture local and global features in HSI. To validate the performance of UDMG, we employ various evaluation methods and conduct experiments on three publicly available hyperspectral datasets. The results demonstrate that UDMG exhibits superior classification performance compared to existing state-of-the-art unsupervised feature extraction methods.

ABSTRACT Thermal infrared remote sensing retrieval is a unique way to acquire large-scale and high-precision land surface temperature (LST), but is often affected by clouds and fog, leading to noticeable spatial gaps. In this article, to address this limitation, we propose a comprehensive framework that integrates satellite and model-simulated data to generate 1 km all-weather LST four times per day across China. Specifically, the framework comprises three key modules. First, we present a data quality optimization preprocessing approach that combines quality flags with morphological processing to balance the quantity and accuracy of satellite LST observations. To address the missing data, a two-stage spatiotemporal fusion method is developed that leverages the complementarity of time-series satellite and model-simulated LST. Additionally, a local-global cascade correction postprocessing strategy is designed to progressively refine the reconstructed results, ultimately achieving stable gapless LST. Upon validation and analysis, the proposed all-weather LST demonstrates a better performance compared to the existing large-scale LST datasets, exhibiting an advantage in mean absolute error ranging from 0.15 K to 0.83 K, verified by in-situ LST. Furthermore, the proposed LST product is consistent with objective geophysical principles and historical meteorological records, which can be anticipated to support research areas such as agriculture and climate change analysis.

ABSTRACT Real-time processing capability is a key enabler for the next generation of intelligent satellites. Full-waveform spaceborne laser altimetry has become an indispensable tool in a wide range of scientific and engineering applications, including terrain mapping, biomass estimation, and Earth system monitoring. Within this context, accurate and efficient laser footprint localization is essential for ensuring the geometric reliability of altimetric measurements. However, conventional waveform-matching methods suffer from severe computational burdens, rendering them unsuitable for high-frequency on-orbit calibration and difficult to deploy in real-time onboard scenarios. To address these challenges, this paper proposes SLA-FLNet – a physics-guided deep learning framework that integrates key physical mechanisms of laser pulse propagation, terrain modulation, and echo formation through a multi-branch spatiotemporal architecture. Each module of SLA-FLNet explicitly encodes a physically interpretable process, enabling accurate, interpretable, and scalable footprint localization. To support supervised training in the absence of ground-truth labels, pseudo-labels were generated using a classical waveform-matching algorithm. The model was evaluated on 2379 laser footprints from 18 beams of the GaoFen-7 (GF-7) satellite, spanning 12 U.S. states with diverse terrain. SLA-FLNet achieved high prediction accuracy and delivered footprint localization results consistent with waveform matching, even in unseen geographic regions. An ablation study further highlighted the critical role of terrain-encoding in enhancing structural fidelity and cross-regional generalization. Compared to traditional methods, SLA-FLNet achieved over 100,000 × inference speedup on modern GPUs, demonstrating strong potential for real-time onboard processing. In summary, SLA-FLNet provides a physically consistent, computationally efficient, and deployment-ready solution for full-waveform footprint localization and on-orbit calibration, supporting future autonomous Earth observation missions.

ABSTRACT With the continuous development and improvement of global navigation satellite systems (GNSS), over a hundred navigation satellites now provide precise and reliable positioning services to various types of users. Most existing satellite selection methods adopt geometric dilution of precision (GDOP) as the selection criterion, which considers only the satellite geometric configuration while neglecting the quality of satellite observations. Consequently, the selected satellite set does not always yield ideal positioning results. To address this issue, we propose a satellite selection criterion that combines the near real-time accuracy of satellite observations with geometric configuration. The near real-time accuracy is estimated based on recent short-term observation data. This method estimates the near real-time accuracy of satellite observations using the Melbourne-Wübbena (MW) combination observation and uses this accuracy to weight and compute the weighted geometric dilution of precision (WGDOP-NRT) for satellite selection. Preliminary experimental results demonstrate that, compared to other criteria, the satellite sets selected using WGDOP-NRT achieve shorter convergence times and higher convergence accuracy in the standard precise point positioning (PPP) model. However, the frequent changes in satellite sets between epochs may negatively affect the positioning performance. To mitigate this issue, we propose an improved satellite selection strategy. If the WGDOP-NRT of the satellite set selected in the previous epoch remains below a predefined threshold in the current epoch, the same satellite set is retained. The method for determining the threshold has also been further discussed. This strategy maintains the geometric configuration and observation quality of the satellite set while enhancing its time correlation in satellite selection. Further experimental results indicate that using WGDOP-NRT and the proposed strategy improves PPP performance in both static and dynamic scenarios.

ABSTRACT Accurate estimation of surface PM2.5 from satellite aerosol optical depth (AOD) is essential for air quality monitoring and exposure assessment. Yet retrieval-based PM2.5 datasets are often biased by the uneven distribution of ground stations, especially in rural areas, causing systematic overestimation of concentrations and related health burdens. To address this, we propose a novel urban-rural balanced estimation framework that mitigates structural bias caused by observation imbalance. The framework integrates AOD, reanalysis fields, and ground measurements within a machine learning model guided by data assimilation principles and further stabilizes performance through ensemble averaging across dynamically weighted background-observation settings. Unlike conventional models that rely solely on ground observations, our approach adaptively constrains non-urban estimates using low-weighted reanalysis samples while preserving high fidelity in urban regions. Applied across China during 2015–2021, the framework reduced root mean square error in clean sites from 18.18 to 8.31 µg/m3, lowered national mean PM2.5 by ~55%, and substantially corrected overestimation of PM2.5-attributable mortality in low-density areas. Results reveal a sharper urban-rural gradient in PM2.5 decline than previously recognized, underscoring persistent exposure disparities. By embedding spatial balance into PM2.5 retrieval and fusion, this study provides a methodologically innovative and scalable pathway to support sustainable development goals on health, climate action, and sustainable cities.

ABSTRACT Changes in forest cover significantly influence carbon-climate dynamics, yet integrated assessment frameworks that connect water-heat-carbon interactions are still limited for subtropical hilly regions. This study presents a spatiotemporal framework for evaluating the impact of forest changes on water-heat-carbon and biophysical processes. Based on effective moving-window samples (Hangzhou: 168; Zhaoqing: 175), we employed partial least squares structural equation modeling (GOF: 0.81/0.71), variance partitioning analysis, and hierarchical partitioning (R 2 : 0.79/0.73) to investigate their interactive relationships and evaluate their accuracy. Results indicate that forest changes exert consistent effects on water-heat-carbon dynamics and biophysical processes in Hangzhou and Zhaoqing, with afforestation and deforestation yielding opposite impacts. The net changes in land surface temperature caused by forest changes from 2000–2010 and 2010–2020 were −0.23 ± 0.40°C/–0.11 ± 0.29°C for Hangzhou and 0.4 ± 0.25°C/–0.49 ± 0.38°C for Zhaoqing, respectively. These changes were significantly larger than the variations in air temperature. Precipitation changes were 13.02 ± 4.66 mm/2.68 ± 5.09 mm and 2.88 ± 1.24 mm/0.62 ± 1.45 mm, with afforestation yielding a greater precipitation increase effect. Carbon stock changes were −8.17 ± 1.79 t/ha/–1.59 ± 1.95 t/ha and −2.76 ± 0.99 t/ha/–2.92 ± 1.28 t/ha. Deforestation resulted in greater carbon losses. We found that the combined effects of biophysical mechanisms exert a greater influence than their individual contributions, and evapotranspiration primarily controls water-heat-carbon interactions. Our study demonstrates that careful management of hilly forests and urban areas is crucial for enhancing carbon sequestration and stabilizing the regional climate. Urban planning should focus on identifying priority afforestation areas. While fully considering the water-heat-carbon benefits of forests, it is also necessary to strengthen forest supervision and establish a monitoring network to achieve sustainable development.