Even though urban growth in metropolitan areas is driven by the interaction between city- and neighborhood-level factors, most conventional deep-learning-based urban growth models rely only on neighborhood-level grid variables, distorting the heterogeneity within and between cities. In addition, their single convolutional kernel learns spatial patterns at a single fixed scale. We propose a multi-modal multi-kernel convolutional long short-term memory (MM-MK-ConvLSTM) model to address the single-scale learning issue in existing deep-learning-based urban growth models. The multi-modal design encodes city-level factors as a city-level context that interacts with neighborhood-level features, enabling the model to learn both levels of factors across the metropolitan area. Furthermore, the attention-based multi-kernels and an encoder–decoder structure learn spatial patterns across multiple scales. Trained on the Seoul Metropolitan Area and applied to Yongin City, the proposed model outperformed ConvLSTM and DS-ConvLSTM across all metrics in late 2010s predictions, reproducing fine-grained fragmented patterns that the other models missed. Moreover, despite having more parameters, the proposed model achieved lower computational operations and faster convergence than both baselines. This demonstrates that multi-scale learning enhances the reliability and efficiency of metropolitan urban growth predictions. Urban expansion to 2040 was simulated using 12 scenarios integrating land demand, slope regulations, and development-scale constraints. Results revealed trade-offs: enhanced slope regulations suppress slope development but increase fragmented development on flat farmlands, whereas combining high-density and development-scale constraints most effectively restricts both slope development and fragmentation. These findings indicate that sustainable land-use planning requires a multi-dimensional policy approach rather than a single regulation.
Resilience of ecological networks (ENs) is a critical indicator of urban ecosystem stability. However, conventional unweighted ecological network assessments deviate from actual ecological conditions by neglecting the spatial heterogeneity of dispersal resistance. Taking the Urban Agglomeration around Hangzhou Bay (UAHB) as a case study, this study constructs a weighted ecological network by integrating species information and embedding dispersal costs into corridor weights. Building upon this foundation, we propose three ecologically modified resilience indicators: the optimal connected subgraph coefficient (S'), the ecological efficiency index (E'), and the ecological resilience index (ERI). These metrics are utilized to characterize the network's response to dynamic disturbances and to conduct a benchmark comparison of eight spatial synergistic optimization strategies. The findings reveal the following: (1) By internalizing species-specific dispersal costs into the weights, we identified 40, 160, 95, 73, and 65 corridors corresponding to levels 1 through 5, respectively. This approach addresses the limitation of unweighted networks, which often overlook local spatial heterogeneity. (2) Under probabilistic and deterministic attack scenarios, the node removal rates when the ERI drops to 0.1 are 0.6 and 0.3, respectively. These values are lower than those derived from traditional resilience indicators. This indicates that the persistence of structural stability does not imply the persistence of functional integrity. Relying solely on structural metrics may underestimate the true vulnerability of the network to functional loss. (3) Regarding spatial synergistic optimization, utilizing our proposed resilience assessment framework, we identified the quantity-priority strategy from a global perspective (G_no) as the optimal approach. This strategy successfully increases the resilience threshold from 0.23 to 0.4, achieving a growth of approximately 74%. This study advances the transition of resilience assessment from geometric connectivity to functional persistence, providing a scientific basis for cross-regional urban ecological governance.
Reservoirs play a crucial role in mitigating floods, alleviating droughts, regulating runoff, and ensuring water security for human use. Accurate quantification of reservoir storage is essential for water resources management, flood control, and regional water security. However, remote sensing-based methods for reservoir storage estimation often assume a horizontal water surface, thereby neglecting the spatial heterogeneity of actual water surface elevations. This assumption can lead to significant errors in reservoir storage estimation during the flood season, particularly for river-type reservoirs with pronounced longitudinal water surface slopes. To address this issue, this study leverages the unique advantage of the Surface Water and Ocean Topography (SWOT) satellite in capturing spatially continuous two-dimensional water surface elevations (WSE). By integrating SWOT data with multi-source remote sensing imagery and digital elevation model (DEM), a quantitative framework is developed to estimate reservoir storage. Using the Gezhouba Reservoir as a case study, this framework enables the generation of high spatiotemporal resolution water surface elevation rasters (WSER) and, through a pixel cloud raster-based reservoir storage estimation approach, achieves reservoir storage quantification from August 2023 to December 2024 while accounting for dynamic storage capacity. Results demonstrate that: (1) The high spatiotemporal resolution WSER constructed in this research accurately captures both the longitudinal variations and spatial heterogeneity of reservoir water levels. Validation against in-situ measurements and the Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) observations reveals strong agreement ([Formula: see text]) and low error (RMSE [Formula: see text] m); (2) Traditional static reservoir capacity estimation methods substantially underestimate reservoir storage during the flood season, particularly under conditions of significant upstream inflow. In contrast, the pixel cloud raster-based reservoir storage estimation method proposed in this study reflects the reservoir storage beneath the actual water surface. The maximum difference in estimated storage between the two methods reaches 0.445 km3, with a maximum relative deviation of up to 69.1%. This research demonstrates the significant potential of SWOT satellite data for regional and even global applications in water resource monitoring, management, and flood risk assessment. It provides a novel technical pathway and methodological framework for the refined and remote sensing-driven management and regulation of reservoirs in the future.
Spatiotemporal interpolation is a fundamental task in geographic information science and a key component of Earth observation data analysis, supporting applications such as meteorological monitoring, air quality assessment, hydrological analysis, and transportation. Incomplete sampling leads to different types of data scarcity. Missing at random (MAR) refers to sparse missing values scattered across observed locations and timestamps, whereas missing completely at location (MCAL) denotes the absence of complete time series at specific locations. Compared with MAR, MCAL is more challenging because there is no temporally contiguous data available for direct interpolation. Existing geostatistical methods provide theoretical interpretability but often lack the capacity to capture complex data structures. Deep learning methods can model nonlinear spatiotemporal patterns, but they commonly suffer from opaque decision-making processes and high computational costs. To address these challenges, this study proposes a Decompositional Kriging Neural Network (DeKNN) for efficient and interpretable MCAL interpolation. The core idea of DeKNN is to shift interpolation from the original high-dimensional spatiotemporal field to a compact latent spatial space. First, a decomposer based on position-aware gated network maps spatiotemporal observations into latent spatial features. Second, a latent-space Kriging neural network estimates latent features at unsampled locations by learning distance-aware covariance structures. Third, a dual-branch reconstructor recovers the complete time series from the interpolated latent features, preserving high accuracy while providing interpretable temporal basis vectors. This decomposition–interpolation–reconstruction design reduces the computational burden of high-dimensional spatiotemporal interpolation while retaining geostatistical interpretability. In addition, the adjoint Kriging mechanism links forward interpolation with backward parameter learning under the same Kriging system. Experiments on two real-world datasets show that DeKNN achieves the best accuracy among the compared baselines and substantially improves computational efficiency. The learned latent spatial features and covariance structures further illustrate the interpretability.
Rapid urbanization has reshaped China’s labor market, but whether policy-led urbanization can improve job quality rather than merely expand employment remains unclear. This study investigates whether and through what channels the New-type Urbanization Pilot Policy (NTUPP) improves urban employment quality in China. We construct an annual panel of 287 prefecture-level cities covering 2007–2021 and exploit the staggered rollout of the NTUPP as a quasi-natural experiment. We estimate policy effects using a staggered difference-in-differences design with city and year fixed effects, supplemented by heterogeneity-robust estimation, dynamic event-study analysis, and a series of robustness tests.The results show that the NTUPP increases urban employment quality by approximately 12.13% on average relative to the counterfactual. Mechanism analyses indicate that the policy effect operates primarily through human-capital accumulation and factor-market development. Urban spatial structure also conditions policy effectiveness: compact development and vertical expansion strengthen the employment-quality gains associated with the NTUPP, whereas excessive horizontal sprawl weakens them. The effects are more pronounced in eastern, southern, coastal, and transportation-hub cities. In addition, the policy is associated with lower intra-city inequality, as reflected in reductions in the urban–rural income gap and the Theil index. These findings suggest that new-type urbanization policies should move beyond population concentration and land expansion. Policymakers should strengthen equal access to education, skills training, and basic public services; deepen labor, land, and capital market reforms; integrate land-use, transport, and industrial planning to promote compact and employment-oriented urban development; and coordinate policy implementation across neighboring cities to reduce spatial inequality and potential displacement effects. Such measures can help sustain inclusive employment gains while advancing decent work and sustainable urban development.
The seeds of the Korean pine (Pinus koraiensis) serve as food for the Amur tiger's prey and also support a multi-billion dollar pinenut market. In northeast China, Korean pine trees are broadly preserved in natural forests and Korean pine plantations have been established across the region. Accurate pine cone yield prediction prior to its harvesting is essential for better planning forest management and preparing for domestic and global pine nut markets. However, traditional pine cone yield estimation relies mainly on tree climbing, which is risky, expensive, and difficult to implement at large scales. Here we developed a UAV imagery-based framework to predict Korean pine cone yield using transfer learning. Two pretrained instance segmentation models, Cascade Mask R-CNN and YOLO11x, were fine-tuned and their performance for cone counting was evaluated across three sites with different tree ages and densities. The results showed that YOLO11x outperformed Cascade Mask R-CNN in cone detection, with AP50 values up to 0.914 for large cones and 0.832 for small cones. In contrast, Cascade Mask R-CNN produced smoother cone boundaries and better mask-level delineation than YOLO11x. Cross-site transferability was proved effective as the introduction of a small proportion of local data (approximately two original UAV images) was sufficient to recover model’s expected performance. Dataset diversity, rather than simply dataset size, is critical for improving model adaptability. Comparisons with field measurements further showed that UAV imagery captured at most approximately 77% of the total cone yield visible from the canopy exterior, and that regression-based correction improved the agreement with field measurements to an R2 of 0.81–0.83. This study also identified an effective UAV acquisition window based on the physiological characteristics of Korean pine and demonstrated the feasibility of UAV-based cone yield estimation. The proposed framework provides a practically efficient solution for pine cone monitoring toward precision management of Korean pine plantations.
River Groundwater Discharge (RGD) is a key component of the hydrological cycle in arid regions, influencing water allocation and ecosystem stability. However, reliably identifying RGD zones at regional scales with satellite thermal infrared remote sensing remains challenging due to limited spatial resolution and unstable detection methods. This study develops a Detection of River Groundwater Discharge (DRGD) approach for identifying potential RGD zones using 30 m high-resolution thermal infrared data from the SDGSAT-1 satellite. The DRGD approach consists of three modules. The first is the Land Surface Temperature (LST) retrieval module, in which LST is derived from SDGSAT-1 data using a split-window algorithm. The second module calculates the Thermal Anomaly Robustness Index (R) using a sliding-window technique constrained by land cover to quantify local thermal anomalies. The third module detects RGD zones by integrating LST, R, topographic wetness index, and normalized difference water index using the Dempster–Shafer evidence fusion technique. The DRGD approach is tested in the middle reaches of the Heihe River Basin. Results show that SDGSAT-1 LST is consistent with MODIS products, with [Formula: see text] and RMSE [Formula: see text] K. A comparison with Landsat 8 LST indicates that SDGSAT-1 preserves finer river-scale thermal variability and more localized low temperature signals. R effectively highlights cooling effects associated with RGD. The multi-source evidence fusion module identified 207 RGD pixels in the study area, achieving a spatial agreement rate of 94.1%, indicating the potential locations of the RGD zones. This study demonstrates the potential of thermal infrared remote sensing for detecting RGD in arid regions. The proposed approach offers a reliable framework supporting hydrological modeling, groundwater assessment, and watershed ecological management.
Understanding vegetation productivity (VP) responses to precipitation variability on China’s Loess Plateau is critical for managing fragile transitional dryland ecosystems, yet these responses and the underlying drivers remain poorly understood. Combining long-term ground and remote sensing observations with an explainable extreme gradient boosting (XGBoost) machine‑learning framework, we found that VP, measured as gross and net primary productivity (GPP and NPP), responded asymmetrically to precipitation anomalies during 2001–2020, with wet-year gains generally exceeding dry-year losses. A five-year moving-window analysis revealed shifting temporal patterns of asymmetry, where wet-year gains peaked during periods of increased precipitation, while drought stress dominated during prolonged dry spells. Asymmetric responses varied across precipitation gradients and biomes, with the magnitude of asymmetry following a unimodal pattern along the precipitation gradient, peaking in semi-arid regions (300–500 mm yr−1). Grasslands exhibited the strongest positive asymmetry due to substantial wet-year gains, while forests showed greater stability and lower asymmetry, reflecting their buffering capacity. Rainfed croplands experienced high sensitivity to precipitation variability but balanced wet- and dry-year responses, resulting in low asymmetry, whereas irrigated croplands displayed reduced sensitivity and higher asymmetry due to irrigation. At the regional scale, VP asymmetry was primarily driven by low-precipitation extremes (31.50%) and interannual precipitation variability (22.86%). However, these drivers varied significantly across biomes: mean precipitation for forests and irrigated croplands, extreme wet‑year precipitation for grasslands, and variability coupled with dryness for rainfed croplands. These findings provide critical insights to convert episodic wet‑year gains into more sustained productivity and improve the resilience and sustainability of dryland ecosystems in the face of increasing climate variability, particularly for predicting regional carbon sink dynamics and formulating adaptive management strategies.
With the arrival of the new breed of very high-resolution (VHR) satellites, which have stereo imaging capability and a spatial resolution of approximately 0.3 m, the extraction of highly accurate digital surface models (DSMs) has emerged as a real alternative to aerial photogrammetry for monitoring complex agricultural and urban environments. This study evaluates the quality of unfilled DSMs derived from Pléiades Neo (PNEO) tri-stereo images over an intensive plastic-covered greenhouse agricultural landscape in Spain. A stereo triplet of WorldView-3 (WV3) was used as the benchmark. The methodology involved sensor orientation and 3D reconstruction using a semi-global matching (SGM) approach. The quality assessment, including both completeness and vertical accuracy metrics, was conducted across three land cover types: plastic greenhouses (GH), urban (U), and bare soil (BS). A highly accurate LiDAR-derived DSM was used as ground truth. A statistical analysis of the quality of the tri-stereo DSMs was conducted using the type of sensor (PNEO and WV3) and land cover (GH, U, and BS) as the main factors. It revealed that land cover was the primary factor influencing DSM quality, explaining 51.72% of the variance in completeness and 77.26% in vertical random error measured as the normalized median absolute deviation (NMAD). Conversely, the sensor factor did not exert a statistically significant influence (p < 0.05) on either completeness or NMAD. Robust estimators, such as NMAD and Median, demonstrated a better ability to deal with outliers and yielded more homogeneous results. NMAD values of 0.32, 0.49, and 0.16 m were achieved for GH, U, and BS, respectively, while completeness ranged from 95.66% to 98.74% for the three different land covers. These findings demonstrate that the new generation of 0.3 m tri-stereo spatial sensors offers an efficient and accurate solution for generating 3D data, even in vertically complex environments. Furthermore, the results have significant implications for plastic pollution management and environmental policymaking.
In Alaska, the presence of permafrost was historically considered a functionally impermeable barrier to environmental contamination. However, climate change in Arctic landscapes has drastically increased the emergence of ecological contaminants from permafrost-bound hazard sites. Here, we showcase the development of ArcGIS Pro 3.x custom geoprocessing tool leveraging synthetic flow modeling and downstream trace analysis for community health decision-making. This toolbox returns highlighted at-risk downstream contamination pathways from synthetic flow models of 5-m IfSAR DTM of local watersheds. We investigated four common Alaskan environmental hazard sites (mines, landfills, active/passive clean-up sites, historic cemeteries) in a case study of tool performance (n = 6 test sites). Double-blind model ground truthing using nearest Euclidean distance GIS validation alongside qualitative validation using satellite imagery shows that tool results strongly agree with real-world hydrology (RMSE 5.8–9.4 m, 13 of 15 ground-truth points successfully routed). This provides support for the utility of this toolbox in real-world water contamination research, mitigation decision-making, and climate adaptation planning. Our Contamination Site Tracing Geoprocessing Tool solves two common problems with native tools, providing the ability to 1) integrate area of interest (AOI) polygons as trace network starting points and 2) run multiple trace networks simultaneously. Customizable features highlight the flexibility of AOI and buffer size to delineate questions of whether contaminant inflow versus terminal outflow tracts are of higher interest to users. Despite work to isolate hazardous infrastructure during development, results show that thawing permafrost-bound contaminants can put thousands of meters of watershed at risk in our use cases. In Alaska, environments function as key subsistence hunting and gathering lands for local rural and indigenous communities. Thus, identifying contamination risk for mitigation planning with limited resources is both a matter of land sovereignty and cultural food security. While our ArcGIS Pro 3.x toolbox and associated workflow documentation were developed for the Arctic, this tool is functionally area-agnostic. Using the tool requires an ArcGIS Pro 3.x series license and access to common data such as area hydrology, digital terrain models (DTM), and researcher-generated AOI polygons. We provide recommendations for tool use with the support of our tested case studies and pathways for tool improvement for users interested in contaminant transport beyond surface flow.
Geographically weighted regression (GWR) provides an important framework for modelling spatial non-stationarity, but its weight-generation mechanism usually relies on isotropic Euclidean distance. In complex geographic environments, physical proximity does not necessarily indicate strong geographic association, especially when adjacent locations differ sharply in spatial structure, directional configuration, or attribute conditions. To address this limitation, this study proposes a context entanglement-based geographically weighted regression (CEGWR) method. Context entanglement is defined as the non-separable coupling among geographic location, directional configuration, and attribute distribution in determining local spatial association. Instead of treating geographic distance and attribute similarity as separate sources of proximity, CEGWR represents each observation through a high-dimensional context vector that jointly encodes spatial, directional, and attribute information. These context vectors are transformed into PCA-whitened context coordinates, and contextual proximity is then used to generate local regression weights. Through this design, CEGWR reconstructs the weighting mechanism of GWR by replacing purely geometric proximity with context-dependent geographic association, while retaining the interpretable local coefficient structure of GWR. In the two simulated settings and an open-source real-world dataset examined here, CEGWR yielded lower fitting errors and AICc values than the compared GWR-type models and improved coefficient-surface recovery in the simulations. These results suggest that, for spatial processes characterized by abrupt transitions or context-dependent local associations, incorporating contextual information beyond geographic distance into the weighting scheme may improve model fitting and the estimation of spatially varying coefficients, providing a flexible local regression framework for analysing complex geographic phenomena.
The TS-DInSAR tool employs a novel and automated approach for the analysis of wide-scale DInSAR datasets acquired via the Sentinel-1 SAR missions and processed using the P-SBAS algorithm. The methodology combines the temporal analysis, employing Principal Component Analysis and K-means algorithms, with the spatial characterization derived from a high-resolution digital elevation model. The initial dataset comprised over than 64 million measurement points, spanning the period from June 2016 to November 2023, and covering the Italian territory. The application of the tool involved the analysis of a selected list of more than 600,000 measurement points, characterized by a mean annual velocity exceeding ±1 cm/yr. The employment of the TS-DInSAR tool ensured the accurate classification of ground deformation phenomena at each measurement point, encompassing landslides, subsidence, tectonic/volcanic activities, uplift, and soil erosion. The development of the TS-DInSAR tool was designed to exploit exclusively time series displacement patterns and spatial parameters, without the requirement of specific inventories, ensuring the reproducibility and scalability of the tool.
Clarifying how Territorial Spatial Conflicts (TSC) evolve across space and time is fundamental to understanding the internal dynamics of territorial spatial systems and to advancing more effective regional spatial governance. Yet existing studies often rely on linear assumptions or correlation-based analyses, making it difficult to capture nonlinear transition processes and disentangle causal relationships among driving forces. Focusing on Yunnan Province, this study develops an integrated analytical framework that combines trajectory diagnostics, interpretable machine learning, and causal inference techniques. TSC evolutionary patterns from 2000 to 2020 were identified at three nested scales—county, township, and 1 km × 1 km grid—distinguishing positive linear, negative linear, positive curvilinear, and negative curvilinear trajectories. On this basis, the LightGBM-SHAP model was employed to uncover nonlinear relationships and threshold effects of “natural-social-economic” factors on the TSC evolutionary trajectories, while the CausalForestDML algorithm was introduced to move beyond correlation and quantify direct causal impacts. The results show that: (1) TSC intensity in Yunnan rose from 0.017 in 2000 to 0.022 in 2020, revealing a composite spatial configuration shaped by “vertical topographic differentiation” and “multi-scale coupling”; (2) Areas exhibiting increasing conflict trajectories accounted for 31.95%, 21.26%, and 17.08% at the county, township, and grid scales, respectively, forming spatially differentiated and multi-centered distribution patterns; (3) Changes in the aggregation index, elevation, and slope emerge as dominant drivers. Their effects are strongly nonlinear and characterized by identifiable thresholds. Elevation and slope exert stable and consistently negative causal influences across trajectory types, whereas change in aggregation index display bidirectional causal effects—reinforcing positive trajectories while mitigating negative ones. By explicitly linking evolutionary pathways with causal mechanisms, this study deepens the understanding of how TSC unfold in mountainous regions of China. The findings provide an empirical foundation for designing spatially differentiated and scale-sensitive governance strategies aimed at balancing development and ecological protection.
Urban land-use mapping is essential for understanding urban functional structures. Point of Interest (POI) data offers advantages for efficient land-use mapping, owing to its lightweight format, accessibility, and socioeconomic semantics. Although recent graph-based models integrate the POI semantics and spatial relationships, they deviate from human cognition due to three main limitations: 1) Node-level semantic objects rarely incorporate fine-grained semantic details, leading to representational ambiguity. 2) Features propagated between semantic objects contain redundancy caused by information overlap among neighboring objects. 3) The inability to model hierarchical “object-functional cluster-land use” relationships. To address these issues, a scale-variant dynamic graph-based framework integrating spatial and hierarchical relationships (DGLU) is proposed for urban land-use mapping. To enhance the discriminative power during node embedding, multi-granular semantic object encoding is proposed by incorporating detailed toponyms with the large language model (LLM). To mitigate feature redundancy in the edge-level spatial relationship modeling, the disparity-aware dynamic graph convolution (DAGC) is proposed to concentrate on distinctive neighborhood features when updating semantic object features. To enable scale-variant hierarchical understanding, DGLU adopts learnable land-use graph pooling (LGP) to integrate homogeneous objects into a “hyper-object”, thereby adaptively generating fine-to-coarse graph structures rather than fixed graph structures. Using open-source areas of interest (AOI) and POI data, we conducted experiments on the POI-CUN dataset covering 34 Chinese cities. The trained model was subsequently applied for urban land-use mapping in the central urban areas of Chengdu, Shanghai, and Wuhan. Both quantitative and visual results demonstrate the effectiveness and superiority of the proposed DGLU framework.
Broadband radiance unfiltering, the spectral restoration of filtered sensor observations, is a fundamental prerequisite for accurate Earth Radiation Budget (ERB) quantification. The newly launched FY-3F Earth Radiation Measurement-II (ERM-II) instrument introduces a dedicated longwave channel, expanding spectral dimensionality beyond its predecessors and necessitating a specialized retrieval framework. This study establishes an operational, scene-dependent unfiltering model for ERM-II that explicitly addresses sensor-specific optical characteristics. A comprehensive spectral library was generated using radiative transfer simulations driven by high-precision global priors. Crucially, to mitigate radiometric distortion caused by sub-pixel scene heterogeneity, we implemented a Point Spread Function (PSF)-weighted convolution scheme. This method integrates collocated high-resolution MERSI-III cloud masks to accurately characterize the effective scene radiance perceived by the ERM-II detector. Based on this framework, optimal unfiltering architectures were determined: a linear fusion of shortwave and total channels for shortwave unfiltering, and distinct single-channel strategies for daytime (linear) and nighttime (nonlinear) longwave retrieval. Validation against Clouds and Earth’s Radiant Energy System (CERES) products demonstrates that the proposed multi-channel fusion effectively compensates for spectral response mismatches. The model achieves a coefficient of determination (R2) exceeding 0.96, with RMSEs constrained below 2.1 W⋅m−2⋅sr−1 and absolute biases within 0.21–1.59 W⋅m−2⋅sr−1. These results represent a substantial quantitative improvement over legacy FY-3C algorithms, reducing systematic biases by up to 96%. This study provides a rigorous physical foundation for the ERM-II mission, preparing it to provide valuable complementary measurements for global climate records in the post-CERES era.
Habitat quality in mountainous Southwest China is shaped by strong terrain gradients, climate heterogeneity, and intensifying human pressures, yet existing remote-sensing ecological indices inadequately represent the productivity and hydrological functions critical to mountain ecosystems. Here, habitat quality is assessed as a remote-sensing-based regional proxy rather than as a direct measure of biodiversity or species-specific habitat suitability. We develop a Productivity–Hydrology Enhanced Remote Sensing Ecological Index (PHE-RSEI) that retains the PCA-based latent-gradient logic of the conventional RSEI while replacing NDVI with EVI and incorporating net primary productivity and water retention to represent productivity and hydrological regulation in mountain ecosystems, and couple it with an integrated diagnostic framework comprising spatial autocorrelation, Theil–Sen/Mann–Kendall trend analysis with FDR-BH correction, Hurst-exponent persistence diagnostics, GeoShapley-based explainable machine learning, and PLS-SEM pathway decomposition. Applied across [Formula: see text] km2 for 2000–2020, PHE-RSEI reveals that habitat quality improved overall: Good [Formula: see text] Excellent area rose from 28.0% to 56.5%. However, the improvement was spatially uneven and potentially unstable: 84.1% of the region showed positive trends, yet only 17.7% were statistically significant, and a mean Hurst exponent of [Formula: see text]0.42 indicates that 67.2% of the improving area faces reversal risk. Driver attribution shows that terrain and climate jointly structure habitat-quality heterogeneity through nonlinear, spatially varying pathways, while human-activity proxies remain a stable net stressor. These findings support persistence-aware ecological zoning that distinguishes fragile high-elevation blocks, stable ecological-function belts, and reversal-prone transition zones, providing actionable guidance for programs such as the Grain-for-Green and Ecological Red Line frameworks.
Surface all-wave net radiation (Rn ) plays a pivotal role in land–atmosphere energy exchange and redistribution, and modulates global water and heat balance and energy circulation. Although direct estimation of Rn from satellite top-of-atmosphere (TOA) observations has shown strong potential, existing methods have largely relied on single-sensor observations and sufficient training samples. This limitation has become increasingly critical as MODIS approaches the end of its mission, while the expected replacement by FY-3D MERSI-II (hereafter MERSI) still lacks enough samples for stable global model development. To address this issue, we first proposed a global Rn estimation framework based on the length ratio of daytime (LRD) for sensors providing TOA observations from the visible to thermal infrared bands. Within this framework, MODIS-only and MERSI-only models were then built separately for daily net radiation ([Formula: see text]) estimation; however, the MERSI-only model showed higher uncertainty due to the limited number of training samples available over a shorter period. To improve the MERSI-only model, we used the MODIS-only model to generate pseudo-labels for numerous unlabeled MERSI pseudo-site samples and combined these pseudo-labeled samples with limited ground-labeled MERSI samples to train a new model using a newly proposed progressive loss-weighting strategy, namely Dynamic Weight Adjustment Training (DWAT). Validation against [Formula: see text] measurements showed that the obtained DWAT model outperformed the MERSI-only model in terms of generalization, robustness, and estimation accuracy, yielding the overall validation and independent-validation root-mean-square-error (RMSE) values of 21.68 and 23.95 W/m2, respectively. The largest improvement was found at high latitudes, with the independent-validation RMSE reduced by up to 3.36 W/m2. Moreover, the DWAT model demonstrated better predictive accuracy and strong spatial mapping ability. Overall, this approach with strong generalization capability provided a practical means to generate accurate and continuous Rn datasets based on limited available samples, and offered a scalable solution for cross-sensor remote sensing applications.
Cloud occlusion is a persistent challenge in optical remote sensing imagery, which substantially degrades the accuracy of semantic segmentation when directly fusing cloudy optical images with SAR data. Conventional approaches typically utilize SAR imagery to aid cloud removal, generating declouded optical images that are subsequently combined with the original SAR data for segmentation. However, these methods often overlook the informative value embedded in the original cloudy features and fail to account for the instability introduced during cloud removal, which may generate unwanted noise detrimental to segmentation performance. To address the above limitation, we introduce PLCSeg, a SAR-guided progressive learning framework for semantic segmentation of cloudy remote sensing images. PLCSeg performs complementary feature fusion and semantic enhancement between cloudy and declouded features across layers from shallow to deep. By fully exploiting both the latent information within the original cloudy imagery and the effective content retained in declouded features, the framework enables end-to-end semantic segmentation under complex cloud conditions. Specifically, in the shallow stage, cross-branch cross-modal feature fusion modules (C-2 FFM) are employed to integrate multiscale features from the SAR, cloudy, and declouded images. In the deep stage, a hybrid mask attention transformer, which comprises mask attention (MA) blocks and adaptive hybrid mask attention (AHMA) blocks, facilitates intra-modal refinement and inter-modal interaction of features. This design effectively suppresses irrelevant regions while emphasizing critical semantic structures. Under the same supervision setting, experiments on the M3R-CR dataset show that PLCSeg improves mIoU and mPA by 1.27% and 0.96%, respectively, over the state-of-the-art. Moreover, it delivers strong performance and robustness across multiple datasets under varying cloud coverage conditions, providing a reliable solution for multimodal semantic segmentation in cloudy scenarios.
Road extraction from high-resolution remote sensing imagery is essential for urban infrastructure monitoring, disaster management, and autonomous navigation. However, this task remains challenging because of limited generalization across domains, topological discontinuity under occlusion, and boundary blurring caused by loss of high-frequency information. To address these issues, this paper proposes a Frequency-Aware Dual-Encoder Network (FADENet) integrating a pre-trained SAM image encoder and D-LinkNet backbone, along with two novel modules: multi-directional Topology-Aware Aggregation (MTAA) for enhancing road connectivity in occluded regions and Adaptive Frequency-Aware Fusion (AFAF) for restoring boundary sharpness. Experiments on the HF road, DeepGlobe, and Massachusetts datasets demonstrate FADENet's superiority in key metrics, validating its effectiveness in improving road extraction completeness and precision in complex urban environments.
