This paper presents the research progress of the author's team in the field of coherent detection infrared imaging technology， explains the sparse aperture infrared imaging method， and analyzes the infrared detection sensitivity for the coherent detection system. It also provides the method for forming infrared complex images with laser local oscillator. For the sparse three-aperture system with sub-mirror structure， the signal processing flow of synthetic aperture imaging is given. In the digital domain， adjacent pupil stitching of sub-mirror is used to reduce the image sidelobes caused by sparse apertures， and a wavelength-tunable laser local oscillator is adopted to reduce the impact of the integration time of direct array detectors on the formation of infrared complex images. The design scheme of a sparse three-aperture prototype system based on direct array detectors is introduced， and the imaging processing results of real data are provided to verify the feasibility of the technology proposed in this paper. Finally， the problems related to array detectors are discussed， and the application of the coherent detection system is prospected.
The identification of technological evolution and channel configuration adaptation patterns for spaceborne conical-scanning microwave radiometers serves as a theoretical foundation for the development of next-generation high-resolution， low-cost radiometers. Taking global spaceborne conical-scanning microwave radiometers as the research object， historical data since the launch of the first instrument in 1974 have been systematically integrated from technical dimensions such as operational lifespan， operating frequency configuration， and scanning observation mechanism， revealing the technological development trajectory of this type of sensor. Spaceborne conical-scanning microwave radiometers have undergone a transformative evolution from single-technology validation to systematic operational application. Core technological advancements are reflected in three aspects： channel configuration has evolved from low-frequency window detection to integrated imaging/sounding capabilities； calibration techniques have advanced from two-point calibration to full-link high-precision calibration； and polarimetric capabilities have progressed from single-polarization to multi-polarization observations. With technological progress， their application scenarios have expanded from early atmospheric and oceanic monitoring to diverse fields such as terrestrial hydrology and polar ice sheet observation. Looking forward， the development of spaceborne conical-scanning microwave radiometers toward higher spatial and temporal resolution is constrained by the dual limitations of physical detection limits and growing application demands. Breakthroughs require continuous innovation in technical pathways such as super-resolution inversion algorithms and constellation collaborative networking， while achieving optimal balance among system performance， engineering complexity， and development cost， thereby enabling coordinated enhancement of observational capability and engineering feasibility.
Against the backdrop of ongoing efforts to strengthen Beijing’s ecological security and enhance ecosystem functions， understanding the spatiotemporal dynamics of ecosystem services and their trade-offs and synergies in Beijing’s Ecological Conservation Area （BECA） has become critical for regional governance and spatial management. Revealing these complex interactions is essential for optimizing ecological spatial patterns and enhancing regional ecosystem functions. This study assessed the spatiotemporal dynamics of six key ecosystem services in BECA from 2000 to 2020 and identified their primary driving factors using the Optimal Parameter-based Geographical Detector （OPGD） model. Pearson correlation analysis and the root mean square deviation method were then employed to quantify trade-offs and synergies among these services. Additionally， bivariate spatial autocorrelation was applied to examine spatial patterns， and a self-organizing map was used to identify ecosystem service bundles and delineate differentiated functional zones. The results show that： （1） From 2000 to 2020， carbon storage and habitat quality slightly declined， while water yield， food supply， soil retention， and recreation services increased. Spatial changes were most pronounced in food supply， recreation， and habitat quality， whereas other services remained relatively stable. （2）The main driving factors included slope， mean annual precipitation， NDVI， and land use intensity. Carbon storage and soil retention were primarily influenced by land use intensity， NDVI， and slope； recreation services were closely associated with nighttime light intensity； and habitat quality was mainly affected by land use intensity， nighttime light， and elevation， exhibiting relatively high explanatory power. （3）Synergies dominated the interactions among ecosystem services， with the strongest synergy observed between carbon storage and habitat quality （correlation coefficients > 0.85）. The most pronounced trade-off occurred between food supply and carbon storage in the early period （correlation coefficients <-0.45）， shifting by 2020 to the strongest trade-off between recreation and habitat quality. （4）Based on ecosystem service bundles， the study area can be categorized into four functional zones： ecological conservation， ecological regulation， agricultural production， and residential development， with corresponding differentiated management strategies proposed. This study provides a scientific basis for refined ecological spatial management and functional zoning in BECA and offers support for ecosystem service-oriented territorial spatial governance.
In recent years， activities such as mountain excavation and land reclamation in Lanzhou City have intensified， which can easily trigger landslide disasters and seriously threaten people's lives and property safety. Therefore， identifying landslide hazard points in the Lanzhou area is of great significance for the early warning and prevention of geological disasters. This paper uses Sentinel-1 data and ALOS-2 data based on Small Baseline Subset Interferometric Synthetic Aperture Radar（SBAS-InSAR） and Differential InSAR（DInSAR） technologies to obtain surface deformation in the Lanzhou area. Combining InSAR observation suitability analysis and surface undulation characteristics， landslide hazard points are identified. Typical threatening landslides are selected for detailed analysis， and the reliability of the verification results of some landslide points through field verification is selected. The research results show that the Lanzhou area can receive backscatter signals for perspective contraction and the proportion of good visibility areas is over 97%. Most areas are suitable for monitoring InSAR deformation information. The low sensitivity area is less than 30%， and most areas have high sensitivity. Synthetic Aperture Radar（SAR） data were used to identify 122 landslide hazard points in the Lanzhou area， among which 79 were identified by Sentinel-1 orbiting tracks and 43 were identified by ALOS-2. Most of the hazard points are concentrated near the Lanzhou urban area along the Yellow River， which is closely related to human engineering activities in the urban area. Field verification results show that the actual deformation features are present in the InSAR deformation areas. Among the 37 identified hazard points， 30 are landslide hazard points， verifying the reliability of the InSAR identification results. Among them， the Lanyaxinghewan landslide near the residential area has an average deformation rate of 72 mm/a per year， and the cumulative deformation is correlated with rainfall； the Weiling Township landslide has been affected by coal mining in this area for a long time， with a deformation rate of 289 mm/a， and the deformation range is gradually expanding， and there are obvious deformation traces on the surface of the slope. After comprehensive comparison， the ALOS-2 data in the L band has a higher recognition rate for landslides in the Lanzhou area than the Sentinel-1 data in the C band. The research results provide important data references for the precise landslide disaster monitoring in the Lanzhou region.
The construction of farmland shelter forests is of vital importance for ensuring high and stable yields in farmlands in Northeast China and safeguarding national food security. Obtaining the spatial distribution of large-scale farmland shelter forests is the basis for analyzing their relationship with productivity. It is of great significance to carry out scientific and reasonable spatial configuration optimization for farmland shelter forests. This paper uses the band reflectance and vegetation index of Sentinel-2A/B images as inputs， and employs the random forest algorithm to select the most suitable feature combination. It then classifies them into land cover categories such as farmland shelter forests， crop land， water， bare， and impervious surfaces. Combined with the land use product， it expands the crop land layer to extract the farmland shelter forests， thereby obtaining the distribution map of farmland shelter forests. The results show that the feature combination consisting of seven features， namely B4， B5， B8， B8A， B12， NDVI， and RENDVI， has the highest classification accuracy. The producer’s accuracy and user’s accuracy of extracting farmland shelter forests are 90.13% and 96.49%， respectively. This paper provides an effective solution for the extraction of large-scale farmland protective forests and the mapping of farmland shelter forests at medium and high spatial resolutions. It can offer data support for macroscopic analysis of the impact of farmland shelter forests on production capacity， and is of great significance for optimizing management and making comprehensive decisions regarding the configuration of farmland shelter forests.
With the launch of the new generation of geostationary meteorological satellites， their high-frequency and high spatiotemporal resolution observation characteristics provide unique advantages for dynamic fire point monitoring. Using AGRI data from China’s new generation geostationary meteorological satellite FY-4A， this study systematically compares and analyzes the monitoring performance and applicability of time-series and context-based methods at different development stages of forest and grassland fires. By calculating the brightness temperature change rate of adjacent times for the same pixel （time-series method） and constructing a local background window to statistically analyze the mean and variance of brightness temperature （context-based method）， the accuracy of fire point identification was analyzed for several typical forest and grassland fires in Sichuan Province and Inner Mongolia Autonomous Region. The results show that in the early stages of a fire， the time-series method is sensitive to sudden changes in brightness and temperature， effectively capturing fire points， with an accuracy rate of 80%~89% for identifying forest and grassland fires. In the middle stages of a fire， the spread of fire leads to significant spatial thermal anomalies， and the context-based method performs better， with an accuracy rate of 82%~88%. In the later stages of a fire， the fire source weakens， and the identification ability of a single method decreases. However， the combined identification strategy of the time-series and context-based methods can significantly improve monitoring capabilities， with the highest accuracy rates reaching 93.56% and 90.36% for forest and grassland fires， respectively. This study confirms that using targeted identification methods for different stages of fire development can effectively reduce missed and false alarms.
As a critical green infrastructure for enhancing urban ecosystem services， accurately acquiring the spatial information of green roofs is paramount for effective urban planning and environmental management. To address the limitations of existing methods in rapidly and accurately extracting green roofs from complex urban built environments， this study proposes a two-stage automated identification framework that integrates deep learning with spectral indices， using Chaoyang District in Beijing as a case study. First， a fully convolutional network featuring an HRNet backbone （FCN-HRNet18s） is deployed to extract building rooftops from 0.8 m High-Resolution GF-2 imagery. Second， green roofs are identified within the delineated building boundaries using the Normalized Difference Vegetation Index（NDVI）.Finally，200 stratified random points are leveraged for validation and accuracy assessment. The results demonstrate that the building extraction stage achieves a pixel accuracy of 0.96 and an F1-score of 0.96， while the green roof identification attains an Overall Accuracy（OA） of 0.82 and an F1-score of 0.80. The total area of green roofs in Chaoyang District totals 598.28 hm2， exhibiting pronounced spatial heterogeneity characterized by a high concentration in a limited number of sub-regions. By successfully combining the spatial feature extraction capabilities of deep learning with the physical merits of spectral analysis， this framework enables high-precision， automated identification within complex urban landscapes， thereby providing robust data support for urban green infrastructure inventories and sustainable municipal planning.
As an important part of air pollution prevention and control and agricultural sustainable development， straw burning control and its precise monitoring technology urgently need breakthroughs. Existing remote sensing monitoring methods mostly focus on the recognition of spectral features of combustion products. However， due to factors such as the variable scale of smoke targets and complex background interference， traditional deep learning models are difficult to achieve an effective balance between detection accuracy and real-time performance. This study proposes a multimodal optimized YOLO v12 improved framework， which significantly improves the detection performance of smoke targets in sentinel remote sensing images by integrating the convolutional block attention mechanism and the multi-scale feature pyramid strategy. The research results indicate that the proposed algorithm， while maintaining a real-time detection speed of 2 ms per frame， demonstrates a superior “accuracy-speed” trade-off characteristic （mAP@0.5 = 95.52%） in the task of straw burning smoke recognition. Based on the XGBoost model for feature importance analysis， the importance of fused features is significantly higher than that of single features， and the vegetation fire-derived index is more effective in extracting combustion information to distinguish targets from interferents. The derived index combination （RGB + MNDFI + BAI + MCRC）， constructed based on the combustion physical process， exhibits better feature separability. This enables the model to achieve an accuracy of 96.39% in identifying the dynamic diffusion process of smoke， and confirms the fundamental principle of remote sensing monitoring： “The richer the spatial details， the stronger the target representation ability.” It is confirmed that medium-to-high-resolution data， with rich spatial details， has a target representation ability that can meet the needs of straw burning monitoring. The spectral collaborative detection framework constructed in this study offers a scalable technical paradigm for the refined supervision of straw burning， and provides a reference for regional air quality warning and the formulation of straw burning ban policies.
Rural roads in China are characterized by large mileage， wide spatial distribution， and frequent location in areas with complex geological conditions. In addition， insufficient maintenance funding makes it difficult to achieve efficient monitoring and prevention using ground-based inspection alone. To address these problems， this study proposes a space-ground integrated method for rural road distress monitoring. First， Sentinel-1A data are used to conduct deformation monitoring with Interferometric Synthetic Aperture Radar （InSAR）， and regional surface deformation risk levels are screened over a wide area. Then， vehicle-mounted ground inspection is carried out to verify and supplement the monitoring results for rural roads with different risk levels. Finally， a space-ground collaborative analysis framework is established. The experimental results show that satellite InSAR technology can rapidly screen deformation risk levels over large rural road areas. Combined with vehicle-mounted ground inspection data， it enables effective point-area collaborative monitoring and supports accurate monitoring of typical road distress sections. The proposed method effectively addresses the limitations of low efficiency and limited coverage in manual inspection， and provides technical support for early identification and precise maintenance of rural road distress.
Ecosystem degradation driven by climate change and human activities has continued to intensify， making ecological restoration a critical pathway for restoring ecosystem functions and supporting regional sustainable development. With the rapid advancement and increasing maturity of remote sensing technologies， their applications in ecological restoration have expanded and deepened. To provide a comprehensive and systematic synthesis of this domain， this study employs bibliometric methods based on the Web of Science （WOS） and China National Knowledge Infrastructure （CNKI） databases to conduct quantitative analyses and an integrative review of 683 SCI-indexed articles and 383 Chinese core journal articles. The results indicate that， internationally， China and the United States are the most prolific and influential contributors and exhibit the closest collaborative ties. International research hotspots primarily concern the restoration of natural ecosystems and mechanisms of response to climate change， whereas domestic research places greater emphasis on artificial ecosystems—particularly mine-site ecological restoration—with monitoring technologies constituting a core research theme. Remote sensing provides high-spatiotemporal-resolution foundational data and enables large-area， continuous monitoring for ecological restoration research. Looking ahead， deeper integration of multi-source remote sensing data is needed to enhance capabilities in key tasks such as identifying ecological degradation， dynamically monitoring restoration processes， and accurately evaluating restoration outcomes， thereby providing robust technical support for fine-grained， full-lifecycle management of ecological restoration.
Red cliffs are the most important identifying element of the Danxia landform and a prominent feature that distinguishes it from other landform types. Multiple spectral transforms （Savitzky-Golay smooth，SG； Continuum Removal，CR； First Derivative，FD； Second Derivative，SD； Reciprocal Logarithm，RL； and Reciprocal Logarithm First Derivative，RLFD） and three characteristic wavelength extraction methods （Successive Projection Algorithm，SPA； Competitive Adaptive Reweighted Sampling，CARS； and Random Frog，RF）， coupled with PLSR-VIP （Partial Least Squares Regression-Variable Importance in Projection） analysis， were employed to extract spectral features from the hyperspectral reflection data of rock faces. We also investigated the relationships between spectral features and mineral compositions of different surface types， as well as the influence of saxicolous plant coverage on spectral responses. Key findings： （1） CARS-optimized PLSR models achieve superior performance， with 19/24 spectral transforms yielding peak test set R²； R² values generally reach 0.99 when RL and RLFD transformations are combined with any feature wavelength extraction method. （2） FD and RLFD transforms can effectively differentiate orange-red surfaces （400~700 nm）， where peak parameter divergences at 440 nm （goethite enrichment） and 570 nm （hematite dominance） reveal distinct mineralogical signatures. （3） RLFD absorption depth variations near 698 nm effectively indicate the degree of vegetation coverage on rock surfaces. The results show machine learning achieves favorable outcomes in hyperspectral feature extraction， improving the efficiency and accuracy of hyperspectral quantitative inversion for Danxia rock hues.
Forest canopy height serves as a crucial parameter for assessing forest structure and ecosystem health， holding significant importance for evaluating forest carbon sinks and analyzing climate change. High-precision inversion prediction of forest canopy data enables potential carbon stock analysis， which is essential for estimating regional forest carbon density and analyzing its dynamic changes. In complex terrain areas， traditional remote sensing methods struggle to obtain accurate data due to topographic interference and vegetation heterogeneity， leading to uncertainties in estimation accuracy. Processed full-waveform features from GEDI spaceborne LiDAR data provide key parameters related to forest structure， including canopy height， vegetation type， and carbon density estimation. Focusing on typical mountainous forests， this study utilized a random forest model with GEDI and multi-source data to construct inversion models for forest canopy height and carbon density. It quantified the relative contributions of various ecological factors to forest structure and， incorporating biological characteristics and meteorological influences， analyzed the spatial differentiation patterns of canopy height and its coupling relationship with carbon sink capacity in typical mountainous regions. Results indicate that the model's inversion prediction achieved canopy height R² values of 0.80 （Yunyang） and 0.63 （Beibei）， with RMSE values of 4.29 m （Yunyang） and 7.18 m （Beibei）， demonstrating strong regional adaptability. Regional forest carbon density ranged from 6.86~147.60 t/hm2（Yunyang） and 3.02-88.56 t/hm2（Beibei）. The continuous distribution characteristics of forests with high canopy closure aligned with the spatial pattern of carbon density. The inversion model exhibited good adaptability across different study areas and quantified the relative contributions of ecological factors， providing scientific basis and data support for forest carbon cycle research and climate change response strategies in mountainous regions. It holds practical value for advancing the dual carbon goals.
High spatiotemporal resolution Land Surface Temperature （LST） plays a direct and important role in studies such as refined estimation of land surface water and heat fluxes， as well as drought monitoring. However， due to limitations such as strip width and orbit height， a single sensor cannot provide LST data with both high spatial and high temporal resolutions. The spatiotemporal fusion of multi-source remote sensing data provides a feasible solution to this problem. Existing traditional spatiotemporal fusion methods typically rely on prior knowledge for model construction and require specific adjustments when applied to different regions， making it difficult to stably reconstruct the required high spatiotemporal resolution of LST. With the rapid development of computer computing power， deep learning models excel in automatic feature extraction and nonlinear modeling. This study takes the Zhangye Oasis Irrigation District and the surrounding areas of Huailai Station as the experimental regions， utilizing two deep learning models： the Enhanced Cross-paired Wavelet Based Spatiotemporal Fusion Networks （ECPW-STFN） and the Multilevel Feature Fusion with Generative Adversarial Network （MLFF-GAN）. Based on Landsat LST and MODIS LST data， high spatiotemporal resolution LST spatiotemporal fusion was conducted， and compared with three traditional methods. The results showed that the mean values of Structural Similarity （SSIM）， Root Mean Square Error （RMSE）， and Mean Absolute Error （MAE） based on deep learning methods were 0.917 6， 1.74 K， and 1.39 K， respectively. The mean values of SSIM， RMSE， and MAE for traditional methods were 0.866 5， 3.22 K， and 2.66 K， respectively. The results based on deep learning methods were overall superior to traditional methods， with the deep learning MLFF-GAN method exhibiting the best performance， with mean values of SSIM， RMSE， and MAE being 0.953 3， 1.51 K， and 1.19 K， respectively. This demonstrates that deep learning models have significant advantages over traditional methods in the spatiotemporal fusion of multi-source remote sensing LST.
Clarifying the spatial heterogeneity of Keyhole reconnaissance satellites imagery at different resolutions during the 1960s—1980s can provide a basis for selecting suitable imagery. This study collected the metadata of Keyhole datasets over China， and classified the imagery into three categories： meter-level （C1）， 5-meter-level （C2）， and 10-meter-level （C3）. The spatial coverage characteristics of Keyhole imagery in China were analyzed. By combining the coverage frequency of free imagery with the cost of paid imagery， the acquisition cost of datasets with triple coverage was estimated. The results show that the coverage proportions of the three categories in the free dataset are 58%， 95%， and 76%， respectively， with average coverage frequencies of 4.9， 4.5， and 3.6. High-value areas are mainly distributed along the Jiangsu-Zhejiang coastal region， the Beijing-Tianjin area， while areas without coverage are concentrated in central-southern Xinjiang and northern Tibet. The complete dataset achieves full coverage across China， with the average coverage frequency increasing by more than an order of magnitude. The spatial heterogeneity of the free imagery is higher than that of the complete dataset. The estimated acquisition costs for triple-coverage datasets of C1， C2， and C3 imagery in China are 750 000 RMB， 750 000 RMB， and 170 000 RMB， respectively. These results indicate that constructing a complete Keyhole imagery dataset covering China to extend the temporal span of earth observation is feasible and can be achieved at relatively low cost.
A systematic calibration framework is proposed for X-band dual-polarization one-dimensional phased array weather radars operating in narrow-transmit-narrow-receive mode. Firstly， static testing is performed on the radar under calibration， encompassing internal/external instrument tests， antenna far-field testing， and metal sphere calibration， to evaluate its performance under near-ideal conditions. Secondly， dynamic testing is conducted， involving precipitation data collection and analysis， as well as assessment of beam discrepancies across different elevation angles. Finally， a consistency comparison algorithm based on multi-range-bin weighting is employed， utilizing the standard S-band weather radar from the Changsha Meteorological Radar Calibration Center for direct performance comparison between the calibrated radar and the reference standard. Test results indicate that the calibrated X-band dual-polarization one-dimensional phased array weather radar generally performs well， with stable beam detection performance across different elevation angles. However， issues such as insufficient sensitivity and lower-than-expected reflectivity factors were observed， necessitating further troubleshooting to ensure compliance with operational requirements.
Optical remote sensing images are often obstructed by clouds， which lead to the loss of critical surface information and substantial degradation of image quality. To overcome this limitation， a cloud removal and reconstruction method for heterogeneous remote sensing images based on MAE-GAN was developed. The proposed method utilizes the cloud-penetrating capability of SAR （Synthetic Aperture Radar） images to complement the missing information in optical images. A Siamese network is employed to extract shared semantic features from SAR and optical images， ensuring consistent representation and robust feature fusion. A masked autoencoder is integrated into the framework to focus on the reconstruction of cloud-covered regions in optical images， while a cross self-attention mechanism is introduced to map SAR features into the optical domain. Adversarial loss and perceptual loss are further incorporated to enhance the recovery of fine details and structural realism， ensuring the reconstructed images are both visually and semantically consistent.Extensive experiments were conducted on the SEN12MS-CR dataset to evaluate the performance of the method. Results indicate that the proposed framework outperforms existing approaches on multiple key metrics， achieving a PSNR of 27.11 dB， SSIM of 0.69， CC of 0.86， MAE of 0.031 49， and RMSE of 0.047 24. Compared to mainstream methods， the proposed method significantly improves the quality of reconstructed images， particularly in terms of texture details， structural authenticity， and semantic consistency. Visual analysis further confirms its ability to restore surface details in heavily clouded regions， making the reconstructed images closely resemble real cloud-free optical images.The proposed method provides a robust and efficient solution for cloud removal in heterogeneous remote sensing imagery， demonstrating significant potential for applications in Earth observation， environmental monitoring， and disaster assessment.
Accuracy assessment of Digital Elevation Models（DEMs） is an important prerequisite for ensuring their reliable application in various fields. Given the difficulty of applying traditional ground-based measurement methods to large-scale DEM accuracy assessment， using data from the second-generation Ice， Cloud， and Land Elevation Satellite （ICESat-2/ATL08） as a reference， combined digital terrain analysis with mathematical statistics to evaluate the quality of eight global DEM products for Anhui Province. The paper also explored the bias characteristics of each product. The results show that FABDEM and GEDTM30 products have the highest overall accuracy， with mean absolute errors of 0.70 m and 0.73 m， respectively. The accuracy of DEM products is significantly affected by topographic factors. For all products， the accuracy is best in high-altitude areas （≥500 m）， high-slope areas （≥25
The integration of drone and handheld LiDAR point cloud data can significantly improve the accuracy and efficiency of forest tree parameter estimation. The Point Cloud Segmentation（PCS）algorithm， which combines region growing and threshold judgment， is used to segment individual trees from UAV LiDAR point clouds. The Density-Based Spatial Clustering of Applications with Noise（DBSCAN）algorithm is applied to segment tree trunks from handheld LiDAR point clouds. The Iterative Closest Point（ICP）algorithm is then used to fuse the point clouds. The individual tree detection rate with fused data reached 96.15%， higher than that of UAV LiDAR alone（93.28%）. For tree height estimation，the fused point cloud achieved an R² of 0.961， RMSE of 0.898 m， MAE of 0.726 m， and MAPE of 7.01%， outperforming both UAV LiDAR（R² = 0.942， RMSE = 0.946 m， MAE = 0.754 m， MAPE = 6.8%）and handheld LiDAR（R²=0.867， RMSE=1.447 m，MAE = 1.224 m， MAPE = 11.6%）. For Diameter at Breast Height（DBH）extraction， the fused point cloud has an R² of 0.979， RMSE of 1.373 cm， MAE of 1.155 cm， and MAPE of 11.87%， comparable to handheld LiDAR（R²=0.967， RMSE=1.286 cm， MAE=1.000 cm， MAPE=10.25%）. In crown width extraction， the fused point cloud achieves an R² of 0.733， RMSE of 0.894 m， MAE of 0.752 m， and MAPE of 35.78%， higher than handheld LiDAR alone（R²=0.561， RMSE=1.131 m， MAE=0.935 m， MAPE=42.42%）. The fused point cloud effectively overcomes the limitations of single data sources， providing an efficient technical means for precise forest resource monitoring and promoting the scientific and refined management of forest resources.
Road extraction from high-resolution remote sensing imagery is often hindered by the presence of clouds， vegetation， and tall buildings， which lead to broken road topologies. Meanwhile， the coupling of multi-scale features significantly reduces extraction accuracy. Existing methods generally suffer from limitations such as semantic discontinuity and loss of fine-grained details when dealing with these complex scenarios.To address these challenges， we propose the Bidirectional Cross-scale Dense Aggregation Network （BCD-LinkNet）. First， we establish an asymmetric dual-path feature interaction mechanism. By integrating a top-down semantic enhancement path with a bottom-up spatial refinement path， we form an encoder-decoder closed-loop information flow that achieves complementary feature fusion across levels. Second， attention-enhanced residual units are introduced into the ResNet34 base module. The Convolutional Block Attention Module （CBAM） is embedded within the architecture and innovatively positioned before residual connections， effectively enhancing the response capability of road features and improving information recall rates. Additionally， a feature refinement strategy is proposed， integrating dual-path feature concatenation with convolutional residual learning mechanisms to enhance feature discriminative power while preserving original spatial resolution. Experimental results on the DeepGlobe dataset demonstrate that the proposed method achieves an Intersection over Union （IoU） of 69.35%， representing a 4.51% improvement over the baseline model， and a Recall of 82.64%， showing a 4.72% increase compared to the baseline. These results validate the effectiveness of the proposed approach in mitigating road extraction discontinuity issues and enhancing extraction completeness.
Accurate estimation of forest aboveground biomass （AGB） is crucial for global carbon cycle research and the implementation of the "dual carbon" strategy. Remote sensing technology has been widely applied in biomass estimation， but research on the spectral-spatial consistency of remote sensing data fusion and model interpretability under complex terrain conditions remains insufficient. Therefore， this paper proposes an improved STARFM （Spatio-Temporal Adaptive Reflectance Fusion Model） image fusion method and a SHAP （SHapley Additive exPlanations） Transformer model， conducting forest AGB estimation research with Qilian Mountain National Park as the study area. The results show that： ① The fused images exhibit excellent performance in spectral-spatial consistency， with an RMSE of 0.021 4 and an AD of 0.005 9， indicating small spectral differences from real images and high overall similarity. ② The Transformer model estimation results reveal that biomass distribution is mainly concentrated in the southeastern part of the study area， and the total forest AGB of Qilian Mountain National Park in 2020 was 2.34×10⁸ t. ③ SHAP analysis indicates that among the global SHAP values， the NDMI， B7， and B5 bands contribute significantly to the model； in the SHAP decision plot， B5_Correlation and B6_Dissimilarity are the main factors influencing AGB estimation. Specifically， B5_Correlation reflects the regularity of forest spatial distribution， B6_Dissimilarity characterizes canopy surface roughness， and there exist complex nonlinear interactions and multi-threshold response characteristics among variables. The image fusion and biomass estimation model constructed in this paper can effectively restore invalid areas in Landsat 8 images affected by cloud cover and shadows， improve the accuracy and interpretability of the model， and provide reliable technical support for subsequent biomass estimation under complex terrain conditions.
Transitional geospatial areas in mountainous regions serve as critical zones for human-land interactions， exhibiting distinct land use evolution mechanisms compared to plains. To uncover its evolution patterns and driving mechanisms， this study focuses on Yongzhou City—a representative transitional geographic space in southern China's mountainous regions. Utilizing Google Earth Engine and random forest algorithms， we constructed a high-precision land use dataset for Yongzhou from 2001 to 2021 （overall accuracy： 91.88%， Kappa coefficient： 0.89）. Results indicate： （1） The overall trend shows “vegetation stability （over 64.5%）-farmland loss （25.67% decrease）-construction land expansion （163% increase）.” Spatial differentiation is strongly constrained by topography： construction land expands radially rather than spreading diffusely. （2） Farmland loss followed dual pathways of “ecological conversion” （43.64% converted to forest/grassland） and “non-agricultural conversion，” highlighting the mountainous region’s response to ecological policies. （3） Combined analysis using Geographic Detector and GBM models （R²=0.866） indicates that natural factors form foundational constraints， while socioeconomic factors serve as core drivers， with nonlinear interactions and threshold effects among all factors. This study reveals the unique coupling mechanism of “topographic constraints-policy responses-economic drivers” in mountainous regions， providing scientific basis for differentiated land management in similar areas.
Assessing the applicability of various spatiotemporal fusion methods in lake dynamic monitoring is of great significance for the optimization of monitoring schemes. Taking the typical area of Poyang Lake as the research object， five mainstream spatiotemporal fusion methods （STARFM， OL-STARFM， FSDAF， OL-FSDAF， and OL-FSDAF 2.0） are selected， and two experimental schemes， namely the dynamic transition group and the water extreme group， are designed. Fusion experiments are conducted in four sub-regions with different geomorphic features， and application validation is performed through water body extraction. The results show that in the dynamic transition group， the highest spectral accuracy of fused images is achieved by FSDAF， while the best performance in structure preservation is obtained by the fused images of OL-FSDAF 2.0. In the water extreme group， robust performance is exhibited by the FSDAF model， which is suitable for processing fragmented water bodies， and outstanding performance in local boundary enhancement is demonstrated by the OL-FSDAF 2.0 model. Water body extraction experiments show that the fused images can generally provide effective data support for lake water body extraction.The synergistic effects of hydrological stages and geomorphic conditions on the performance of spatiotemporal fusion models are revealed， and the deficiencies of existing literature regarding the application of spatiotemporal fusion methods in dynamic environments are remedied. Theoretical support is provided for the selection of optimal fusion models under different monitoring scenarios. Multi-source spatiotemporal fusion images effectively compensate for the limitations of single data sources， and high-frequency and high-precision data support is supplied for water resource management， ecological protection， and flood monitoring.
