In this study， a comprehensive overview of snow avalanche hazards occurred in the past in Xizang area was reviewed first， followed in-depth analysis on the spatial distribution and main driving factors of snow avalanche hazards in the Xizang mountain region. Snow avalanche-prone areas for the study area were then mapped based on the spatial distribution of snow cover and DEM （Digital Elevation Model） data， and were validated using in-situ observations in southeastern Xizang. Results indicated that there are the highest frequencies of avalanche occurrences in southeastern Nyainqentanglha mountains and southern slope of the Himalayas. In the interior of plateau， avalanche occurrence is constrained due to less precipitation and flatter terrain. The perennially snow avalanche-prone areas in Xizang account for 1.6% of total area of the plateau， while it reaches 2.9% and 4.9% of total area of Xizang in winter and spring， respectively. Snow avalanche hazards and fatalities present increasing trends under global climate warming due to more human activities at higher altitudes. In addition to continuous implementation of engineering prevention and control measures in the key regions， such as in Sichuan-Xizang highway and railway sections， enhancing monitoring， early warning and forecasting services are important to prevent and mitigate avalanche hazards in the Xizang high mountain regions.
Monitoring of glacial lakes based on remote sensing is one of the current research hotpots in cryospheric science. Imageries from the Multispectral Imager for Inshore of SDGSAT⁃1 would be one of important data resource for glacial lake researches. Our study focuses on nearly 2 000 glacial lakes in the Mount Everest region. An automatic extraction for glacial lakes is applied to the multispectral bands of SDGSAT-1 images. The accuracy of the glacial lakes on SDGSAT-1 MII images is assessed by comparing with lakes from both Sentinel-2 MSI and Landsat 8 OLI. The area difference of glacial lakes between SDGSAT-1 and Sentinel-2 MSI or Landsat 8 OLI is quantified as error. And the Jaccard coefficient performances the ratio of overlapping between glacial lakes from SDGSAT-1 and Sentinel-2 MSI， which indicates the position bias between two lakes. Coregistration error between bands 2 and 7 of SDGSAT-1， band 2 of SDGSAT-1 and band 3 of Sentinel-2 is calculated by COSI-Corr software. And the correlation analysis between co⁃registration error and area error， Jaccard coefficient is employed in SPSS software. The results indicate that the decision tree based on Normalized Difference Water Index enables rapid extraction of glacial lake vector data from SDGSAT-1 multispectral imagery. The accuracy of glacial lakes on SDGSAT-1 is higher than those from Landsat 8 OLI but lower than those from Sentinel-2 MSI， with an average error of 16.6%. Glacial lakes in shadowed areas on imagery of SDGSAT-1 after atmospheric correction was more exactly outlined than that before， although there was 10% more error compared to other glacial lakes. The area error of glacial lake on SDGSAT-1 is mainly defined by co-registration error of bands 2 and 7 of SDGSAT-1， while the Jaccard coefficient is influenced by the error from both bands 2 and 7 of SDGSAT-1， and band 2 of SDGSAT-1 and band 3 of Sentinel-2. SDGSAT-1 multispectral data serves as a robust data source for studying glacial lakes. Although its accuracy of identifying supraglacial lakes on SDGSAT-1 MII is lower than for other lakes， SDGSAT-1 MII will be the potential data for researches of the identification of supraglacial lakes and the change of supraglacial lakes.
Snow plays a crucial regulatory role in the ecosystem of the Tibetan Plateau and has a profound impact on carbon storage and the carbon cycle. As a key variable in the carbon cycle， the spatiotemporal variation characteristics of the Gross Primary Productivity （GPP） of surface vegetation and its response mechanism to snow parameters are directly related to the carbon budget of alpine grassland ecosystems. However， current research on the impact of snow changes in the Tibetan Plateau on GPP is still insufficient， especially the regulatory role of snow phenology （such as the start and end dates of snow cover） on GPP has not been clarified. This study， based on multi-source remote sensing data including the GPP dataset from 2000 to 2018 and snow phenology data， analyzed the spatiotemporal distribution characteristics of GPP and snow parameters in alpine grasslands of the Tibetan Plateau. Combined with meteorological data such as near-surface air temperature， downward shortwave radiation at the surface， and precipitation， the study explored the influence mechanism of snow parameters and meteorological factors on GPP in alpine grasslands. The results show that the annual average GPP value in the Tibetan Plateau presents a distribution pattern of being lower in the northwest and higher in the southeast， and shows an overall increasing trend year by year. Both the start date and end date of snow cover show an advancing trend， while the duration of snow cover shows a decreasing trend. The annual average GPP increases with the delay of the start date of snow cover， the advance of the end date of snow cover， and the reduction of the duration of snow cover. The dominant factors influencing the change of GPP are， in order， near-surface air temperature， duration of snow cover， precipitation， downward shortwave radiation at the surface， start date of snow cover， and end date of snow cover. Although meteorological factors have a greater impact on GPP， the influence of snow parameters on GPP cannot be ignored， especially on spring GPP. The change of GPP is comprehensively affected by the spatial distribution of snow parameters and meteorological factors as well as their interaction. This study reveals the regulatory role of snow dynamics on GPP in alpine grasslands and clarifies the relative contributions of snow parameters and meteorological factors to GPP. The research results are helpful to deepen the understanding of the carbon cycle process in alpine ecosystems of the Tibetan Plateau and provide a scientific basis for predicting the dynamic changes of GPP under the background of climate change.
Fractional Snow Cover （FSC） is a key parameter to measure the distribution of surface snow cover. Deep learning has become an important method to estimate FSC， and the quality and representativeness of training samples directly affect the accuracy and generalization ability of the deep learning model. This study takes Xinjiang as the research area， the performances of four representative sample selection methods， namely hierarchical random sampling， cosine similarity， K-means clustering and information entropy， were compared during the process of MODIS FSC mapping by using deep learning methods. The results show that the four representative sample selection methods can effectively screen out representative samples， reduce sample redundancy， and improve the estimation accuracy and efficiency of FSC deep learning estimation models. Compared with deep learning models trained with a total sample， the model constructed based on the representative samples screened out by K-means clustering （accounting for approximately 40% of the total sample） performs the best in the FSC estimation accuracy. The RMSE and MAE were 9.51% and 5.272% respectively， and R² has reached 0.916， and reduce the calculation time by about half. In conclusion， the representative sample selection strategy can significantly enhance the applicability of the deep learning FSC estimation model in complex environments， providing a feasible path for its transformation from experimental research to operational applications.
Accurate inversion of Fractional Snow Cover （FSC） in forested areas is significant for hydrological process simulation， climate change projection and ecosystem management. This study proposed a hybrid machine learning model， RF_ART， based on the Random Forests （RF） algorithm and the Asymptotic Radiative Transfer （ART） model， aiming to improve the inversion accuracy of FSC in forested areas. The model integrated multiple environmental variables， including spectral characteristics， vegetation， terrain， angles， land surface temperature， and snow grain size. Experiments were conducted in the Altay region and the central-eastern Tianshan Mountains of Northern Xinjiang. The results showed that the average Root Mean Square Error （RMSE） of RF_ART in the training images was approximately 0.048 0， and the average RMSE in the testing images was approximately 0.096 6， which was significantly lower than those of the NDSI_FSC and NDFSI_FSC methods. Additionally， while the RMSEs of RF_ART and RF_FSC were similar， in testing images， RF_ART introduced physical constraints that enhanced the model's robustness， making it the preferred algorithm for FSC inversion in forested areas. Notably， in the predominantly deciduous forests， the RMSE of the RF_ART model gradually decreased with the incorporation of various variables. Moreover， under conditions of scarce data， the RF_ART model demonstrated strong robustness and application potential. By combining hybrid machine learning models with multi-source remote sensing data， this study provides an important reference for the inversion of FSC in forested areas.
Owing to complex climatic and topographic constraints， optical and microwave remote sensing applications in the Meili Snow Mountains region face significant limitations. This study employed high-precision photogrammetry using a DJI Mavic 3E drone equipped with a Haixingda iRTK20 GNSS system to survey the terminus of the Gongsenglongba Glacier on October 1， 2023， and October 1， 2024. Automated extraction of glacial morphological features was accomplished using a U-Net deep learning model. The result shows： The U-Net model achieved high accuracy and exhibited strong generalization capabilities in extracting glacial morphological features. Key observations during the study period （October 2023 to October 2024） include： a significant average surface elevation reduction of 2.37 meters at the glacier terminus； a continuous expansion trend in supraglacial lakes with relatively uniform spatial distribution； widespread development of ice cliffs showing progressive expansion towards the direction of glacier retreat， particularly pronounced in the middle and upper terminus sectors； and a concentrated distribution of supraglacial crevasses within the 3 910~4 000 m and 4 110~4 200 m elevation bands， contrasting with sparse occurrence elsewhere. This research establishes an intelligent extraction framework for glacial morphological features by integrating drone imagery with deep learning methodologies. The developed approach provides robust technical support for high-precision monitoring of glacial dynamics and their climatic responses.
Snow cover monitoring is critical for climate change studies. While UAV-based RGB imagery enables rapid data collection， it often fails to differentiate snow from spectrally similar surfaces in complex terrain. This study integrates visible-thermal infrared imagery with ensemble learning， evaluating 10 classifiers' performance across full and optimized feature datasets to systematically improve snow identification accuracy and operational efficiency. The results indicate that： （1） Seven ensemble learning classifiers exhibit significant advantages in snow cover monitoring. For example， the AdaBoost-DT classifier achieves an Overall Accuracy （OA） of 90.54%~92.91%， outperforming single strong classifiers （e.g.， CART， SVM， and MLPC） by approximately 1%~6% in OA； （2） Utilizing full feature datasets retains more feature information， improving OA by 1%~3% in complex snow distribution scenarios. Optimal feature selection reduces redundant features and shortens classification time， but may lead to notable accuracy degradation for certain classifiers （e.g.， LGBM and GBDT）； （3） Key features such as Mean， Blue Band （B）， Thermal Infrared Band （T）， and Red-Blue Difference Index （RBDI） substantially contribute to snow cover monitoring accuracy， serving as critical indicators for snow cover identification.
Accurate extraction of the active layer thickness in permafrost is essential for advancing the understanding of permafrost degradation and its environmental impacts. Ground-Penetrating Radar （GPR） has become a key method in this research due to its efficiency， non-destructive nature， and high accuracy. However， traditional manual interpretation is labor-intensive， time-consuming， and susceptible to data loss.To address these limitations， this study pioneers the application of a correlation-based horizon tracking method for automated interpretation of GPR images and extraction of active layer thickness in a typical permafrost region of the Qinghai-Tibet Plateau. By optimizing seed point identification and tracking， the precision of active layer thickness extraction was further improved.Using GPR data from 26 survey lines in the Qilian Mountains permafrost area， the active layer thickness was determined using both manual interpretation and the horizon tracking method. The results demonstrate that the improved horizon tracking method exhibits lower errors compared to the original approach， with reductions of 0.025 m in Bias， 0.018 in RE， and 0.039 in RMSE. Additionally， it shows a stronger correlation with manual interpretation results（improved by 0.014）， and the main portion of the GPR image aligns more closely with manual interpretations. These findings confirm the reliability of this method for studying permafrost active layers.
As the third pole of the Earth， the soil freeze-thaw cycle plays a crucial role in climate change research over the Tibetan Plateau. In order to explore the technical methods for identifying the soil freeze-thaw status and revealing the intra-annual cycle characteristics of the soil freeze-thaw process in the alpine meadow of the source area of the Yellow River， the observation data from the Soil Moisture and Soil Temperature observation networks （SMST） and the ground-based dual-polarization microwave L-band radiometer in the source area of the Yellow River are deployed in this study. The soil freeze-thaw status are identified by the traditional method and the Freezing Factor （FF）， and the intra-year characteristics of the soil freeze-thaw process characterized by the two datasets were analyzed， and the results were verified each other and discussed. The results show that the soil freeze-thaw status monitoring by ground-based microwave radiometer observation data are basically consistent with the soil freeze-thaw process characterized by top-layer soil temperature and moisture. Of the four freezing factor methods， the Polarization Ratio Frost Factor （FFPR） and the Combined Dual Polarization Difference Frost Factor （FFCDPD） are the most applicable for identifying soil freeze-thaw states. When the incidence angle of the ground-based microwave radiometer is at 50°， FFCDPD can be chosen as the optimal one for identifying the soil freeze-thaw status， and the accuracy is 84.2% during the whole annual cycle period； When the incidence angle is at 40° or 60°， FFPR can be regarded as the optimal frost factor， and the accuracy are 91.4% and 91.2%. When the soil freeze-thaw status is transferring from completely thaw to freeze or freeze to thaw， the bright temperature under vertical polarization and horizontal polarization decreases or rises at the same time because the quickly change of the liquid water content， and the polarization brightness temperature difference is between 40.0 K and 75.0 K. Therefore， the freeze-thaw state of the top-layer soil can be effectively identified by using ground-based microwave radiometer observation data and different frost factors. The results of this study can provide a scientific reference for satellite microwave remote sensing to identify the soil freeze-thaw and water cycle over the alpine meadow.
Evapotranspiration （ET） plays a critical role in linking the carbon， water， and energy cycles. Accurate estimation of ET is essential for understanding ecological and hydrological processes in karst regions and provides crucial support for managing water resources and ecological environments. Observational data from ET monitoring stations and atmospheric reanalysis datasets were employed to develop and optimize four machine learning models （e.g.， Support Vector Machine （SVM）， Long Short-Term Memory （LSTM）， Backpropagation neural network （BP）， and Random Forest （RF）） using the Particle Swarm Optimization （PSO） algorithm. Model accuracy was evaluated through both random and spatial cross-validation frameworks. Additionally， model interpretability was assessed using SHapley Additive exPlanations （SHAP） to identify the most influential input factors. The optimal model was then applied to simulate ET across the peak cluster depression basin in southwestern Guangxi. The results reveal the following： （1） Under random cross-validation， the PSO-SVM model demonstrated the highest accuracy， while under spatial cross-validation， the PSO-LSTM model performed best. The respective mean values of correlation coefficient （R）， coefficient of determination （R2）， and Mean Absolute Percentage Error （MAPE） were 0.941 0 mm·m-1， 0.869 1 mm·m-1， and 0.092 8 mm·m-1 for PSO-SVM， and 0.919 7 mm·m-1， 0.796 5 mm·m-1， and 0.120 6 mm·m-1 for PSO-LSTM. （2） According to SHAP analysis， net surface solar radiation emerged as the most influential factor in the PSO-SVM， PSO-BP， and PSO-RF models. In contrast， precipitation was identified as the most significant driver in the PSO-LSTM model. All four models consistently captured temperature as a key factor affecting ET. （3） The annual average ET in the study area exhibits an overall increasing trend. High ET values are concentrated in agricultural zones of Chongzuo due to irrigation， whereas regions with lower ET values correspond with the distribution of karst landforms. This study demonstrates that combining ground-based meteorological observations， remote sensing data， optimization algorithms， and machine learning techniques can achieve high-precision， large-scale ET estimation in complex karst environments.
Forest aboveground biomass （AGB） is a crucial indicator for measuring forest carbon sinks and assessing the carbon sequestration capacity of forests. The Global Ecosystem Dynamics Investigation （GEDI）， a spaceborne LiDAR system specifically designed for forestry， plays a significant role in providing high-precision AGB data. However， there is currently a lack of systematic accuracy assessments for GEDI L4A products， and it remains unclear which factors most significantly influence their accuracy. This study utilizes an airborne AGB dataset to validate the AGB accuracy of GEDI L4A and analyzes the effects of topography， GEDI quality indi-cators， and other environmental factors on AGB estimation accuracy. The main conclusions are as follows： （1） Through the accuracy analysis of the study areas， the Root Mean Square Errors （RMSE） for Hood Canal， King County， and Solduc are 124.76， 146.96， and 154.11 Mg/hm2， respectively， indicating that the accuracy in Solduc is relatively poor. （2） The primary factor affecting AGB estimation accuracy is topography， followed by the Signal-to-Noise Ratio （SNR）. （3） The accuracy of AGB estimates decreases as slope increases； sunny slopes have better estimation accuracy than shaded slopes； and as the signal-to-noise ratio increases， the accuracy of AGB estimation also improves.This study validates the accuracy of the L4A data product and provides an in-depth analysis of how various influencing factors affect AGB estimation accuracy， offering important references for selecting high-quality GEDI biomass samples.
Leaf Area Index （LAI） is an important parameter reflecting crop growth. Its accurate acquisition is crucial for agricultural monitoring and yield assessment. The Sentinel-2 satellite has multiple red-edge and shortwave infrared bands， which have potential advantages in LAI estimation. Therefore， comparing the estimation capabilities of different models and band combinations is of great significance to improving the accuracy of maize LAI estimation. This study uses the Daman Observation Field at the Heihe Remote Sensing Experimental Station in Zhangye， Gansu Province， as the research area. Based on PROSAIL model sensitivity analysis， we screened band combinations and key parameters sensitive to LAI and constructed a simulation database. LAI was inverted using three methods： a Look Up Table （LUT）， a Genetic Algorithm （GA）， and a Random Forest （RF）. Accuracy was verified using Sentinel-2 imagery and field data. The results show that： ①outliers significantly affect estimation accuracy. LUT is most sensitive to outliers （ΔR2=0.20~0.26）， while RF is relatively stable （ΔR2=0.14~0.20）. Adding the red-edge band RE2 （B， R， RE2， RE3， NIR， RE， SW2） to the LUT improves accuracy while maintaining interference resistance （mean ΔR2=0.18）. ②After removing outliers， LUT achieves the highest inversion accuracy （R2 = 0.88， RMSE = 0.31）， followed by GA. Adding RF to the RE2 band significantly improves performance （R2 = 0.65~0.79， RMSE = 0.64~0.53）. ③The inversion accuracy of the three models in the high LAI range （2.5~5.0； LUT： R2>0.84； GA： R2>0.71； RF： R2>0.57） is significantly better than that in the low LAI range （0.5~2.5）， among which the R2 of RF combined with RE2 band can increase from 0.57~0.82. In summary， the physical model inversion method and RE2 band play an outstanding role in improving the accuracy of maize LAI estimation， and can provide some valuable references for maize LAI estimation and growth monitoring.
Complete urban surface temperature （Tc） is of great significance for the study of urban microclimate， urban energy balance， and urban surface heat island. There are currently two methods for estimating Tc using remotely sensed directional radiometric temperature including the regression model method and the integra， but their applicability and accuracy have not been verified. Based on the simulation data， 10 types of representative Urban Structure Models （USM） were constructed based on Local Climate Zone. The accuracy of the two methods is compared from the Angle of building structure and component temperature difference. By exploring the ability of different estimation methods to simulate the temporal and spatial heterogeneity of Tc， the applicability of the model was analyzed. The results show that： ① The average root Mean Square Errors （RMSE） of SIT and RMT are 1.4 K and 1.1 K， respectively. In terms of estimation accuracy， the large horizontal surface area is more favorable for RMT simulations. Therefore， the accuracy of RMT is higher whereas the accuracy of SIT is lower in urban scenes with smaller F （the ratio of total building wall area to the gross area of the lot）. Specifically， RMT is applicable in USM01 or USM05~10 whereas SIT is applicable in USM02~04. In terms of stability， RMT is based on global data whereas SIT is based on local data， so solar radiation， building materials and meteorological condition have a higher impact on RMT than SIT. ② Among the four temperature types （Tc， SIT， RMT and Tnadir）， Tc presents the highest spatial heterogeneity， whereas Tnadir presents the largest diurnal and annual variations. SIT outperforms RMT in depicting the spatial heterogeneity of surface temperature. RMT outperforms SIT in simulating the temporal variation of Tc during nighttime whereas the opposite is true during daytime.
The rapid industrialization has led to increased mining activities and exacerbated environmental pollution， making bare soil identification crucial for monitoring the geographical distribution and spatial changes in mining areas. Traditional methods for bare soil identification rely on medium-resolution imagery， often facing challenges related to algorithmic efficiency and accuracy. To address this， we propose a BareFormer model that combines superpixel technology with Transformer architecture. This method integrates a superpixel spatial self-attention module， enhancing the learning capability of local texture features and long-range dependencies. A bare soil dataset containing 3 759 pairs of high-resolution mining area images and labels was created to test the classification performance of different methods for fine-grained bare soil in mining quarries. The study results indicate： ① BareFormer achieves superior performance in bare soil classification， with an accuracy， recall， F1 score， and IoU of 87.30%， 88.30%， 87.79%， and 79.77%， respectively. ② Ablation studies and model complexity experiments demonstrate the effectiveness of the superpixel attention mechanism and multi-scale feature fusion in improving model performance. ③ Activation map analysis and robustness tests show BareFormer's strong capability in capturing bare soil details. These findings provide an efficient and accurate new method for the fine-grained identification of bare soil in mining areas， with significant practical application value and research implications.
The study of spatiotemporal changes along the coastline has important guiding significance for coastal ecological environment protection and resource planning. Based on the GEE platform， Landsat remote sensing images from 2000 to 2023 were screened to calculate mNDWI（Modified Normalized Difference Water Index）， EVI（Enhanced Vegetation Index）， NDVI indices（Normalized Difference Vegetation Index） for obtaining water edge frequencies. The water edge frequency was calculated and visually interpreted to compare with high-precision remote sensing images to select segmentation thresholds for extracting the coastline of the Jiaodong Peninsula. This article uses the DSAS system（Digital Shoreline Analysis System） to analyze the spatiotemporal variation patterns of coastlines and show that from 2000 to 2023， the overall length of the coastline of the Jiaodong Peninsula increased by 285.77 km， with an increase of 12.42 km/a. The coastline advanced towards the sea by 126.58 m， with a rate of 8.53 m/a； Taking 2005 as the node， the length and movement rate of the coastline from 2000 to 2005 were lower than the average level. After 2005， the coastline grew rapidly and moved faster towards the sea. Among them， the growth rate of coastline length from 2010 to 2015 was the highest， reaching 47.55 km/a； The speed of coastline movement towards the sea has also accelerated rapidly since 2005， with the period from 2005 to 2010 being the fastest period of movement，NSM 69.56 m，LRR 14.81 m/a. On a spatial scale， the coastline of Yantai City has the greatest degree of expansion， NSM 212.81 m，LRR 10.92 m/a； The NSM and LRR of the coastline in Weihai City are the lowest， and the degree of change is the smallest； The changes in Qingdao's coastline are basically on par with the average level of the entire region. Due to the fact that coastal cities and coasts are mainly affected by human activities， an analysis of the intensity of human activities along the coast of the Jiaodong Peninsula found that since 2005， the area of high-intensity human activities in the coastal region of the Jiaodong Peninsula has significantly increased. It is preliminarily believed that the coastline of Jiaodong Peninsula pushed towards the sea as a whole from 2000 to 2023， and 2005 was the key node of change，and then the speed accelerated significantly，and human activities were the main driving factors.
This study investigates the utilization status of arable land resources in Jilin Province and provides theoretical support for promoting their sustainable development. The Per Capita Cultivated Land Area Model and EF-NPP Model were used to assess the ecological balance and supply-demand dynamics of arable land in Jilin Province from 2001 to 2021. The Sustainable Utilization Index and Grey Prediction Model were then employed to predict the sustainable utilization trend of cultivated land from 2022 to 2031. Finally， rational control zones were proposed based on the findings.（1） From the perspective of food security， the supply-demand profit and loss coefficient of arable land increased from 0.532 to 0.710， indicating a trend of “surplus orientation and gradual optimization”.Spatially， this trend shows a “surplus in the central and western regions and a deficit in the southeastern region”.（2） From the ecological security perspective， the ecological profit and loss coefficient of cultivated land improved from -0.201 to -0.118， reflecting a positive trend. Spatially， this pattern reveals a “loss in the west and surplus in the east”. The Sustainable Utilization Index increased from 0.455 to 0.472， with projections indicating that cultivated land will remain in a weakly unsustainable utilization state （Ⅲa） from 2022 to 2031， suggesting that the sustainable utilization level has not yet reached an ideal state.（3） Based on the comprehensive profit and loss value and relevant policies， the utilization of arable land resources is classified into three types for zoning control： stable improvement， potential development， and rectification optimization. This classification aims to enhance the level of sustainable utilization and promote the sustainable development of arable land. The proposed control zones are expected to contribute to achieving sustainable utilization goals.
Small unmanned vessels can measure continuous underwater topography and provide important and key data for remote sensing runoff monitoring. Yet their monitoring effects in rivers of different forms are still unclear. This study comparatively analyzes the accuracy differences in remote sensing-derived runoff calculations using USVs across diverse river systems. The results show： ①In four sites where the river width is greater than 40 m， the accuracy improvement effect of the unmanned boat measurement method on the shape of the river cross-section is more obvious； ②Compared with the data of the national hydrological station， the average values of NSE coefficient and RMSE coefficient for runoff calculated by the traditional cross-section measurement method are 0.60 and 3.57 m³/s respectively， while the average values of NSE coefficient and RMSE coefficient for runoff calculated by the unmanned boat measurement method are 0.81 and 3.22 m³/s respectively；③ The average range of absolute errors in runoff calculated by the traditional cross-section measurement method is 1.82~12.76 m³/s， which is greater than the average range of 0.95~8.57 m³/s calculated by the unmanned boat measurement method. By capturing high-resolution bathymetric data to refine cross-sectional profiles， the Small unmanned vessel approach enhances runoff calculation accuracy and is particularly effective in rivers with surface widths greater than 40 m.
In the context of global climate warming， exploring the impact of winter climate variability on vegetation growth is beneficial for understanding the role of the non-growing season in the annual land-atmosphere interaction changes. We utilized GEE cloud platform to acquire GIMMS NDVI data （1985~2015） and employing Sen's trend analysis and the coefficient of variation method， this study analyzed vegetation dynamics in Northeast China from 1985 to 2015. Combining winter climate data （minimum temperature （Tmin）， maximum temperature （Tmax）， and precipitation （Prec））， the study used Spearman's correlation analysis to examine the response of vegetation NDVI changes to winter climate. The results show that from 1985 to 2015， the annual vegetation NDVI variation in Northeast China shows a fluctuating increase at a rate of 0.0012/a， with the most significant trend observed in coniferous forests， increasing at a rate of 0.006/a. The average coefficient of variation for vegetation NDVI in Northeast China from 1985 to 2015 is 0.05， indicating overall stability， with the highest coefficient of variation， 0.74 observed in the Hulunbuir grassland in the west. Winter temperatures （Tmin and Tmax） shows an increasing trend， while precipitation generally shows a decreasing trend. Vegetation NDVI in Northeast China is negatively correlated with winter precipitation （Prec） and positively correlated with winter Tmin and Tmax， with clear spatial characteristics. The results of this study help to understand the mechanisms of winter climate variation and vegetation response in Northeast China and provide a reference for the assessment and management of the ecological environment in the region.
Using radar remote sensing technology for large-scale identification and monitoring of landslide-prone areas is a current focus in disaster prevention and mitigation. High-voltage power transmission lines， as an essential part of the national power grid， can suffer significant economic losses if landslides occur in their vicinity. This study employs SBAS-InSAR technology and optical remote sensing imagery to conduct an in-depth investigation of landslide hazards and influencing factors in Longyan City， Fujian Province， with particular attention to landslide-prone areas near power lines. The research findings indicate that landslide disasters in Longyan City predominantly occur in areas with slopes and steep inclines， at elevations ranging from 200 m to 600 m， and are closely related to seasonal rainfall and mining activities. In addition， this study focused on seven significant deformation areas located within two kilometers of the transmission towers， with annual deformation rates ranging from -0.25 to -0.6 m， posing a direct threat to the transmission infrastructure. These research findings have been validated through field investigations， confirming the critical role of the proposed method in ensuring the safe operation of transmission lines.
As a resource-transition city， Xuzhou’s estimation of ecosystem carbon sequestration capability is critical for achieving the "dual carbon" goals. The Light Use Efficiency （LUE） model plays an important role in assessing the carbon fixation ability of regional vegetation. In this study， the LUE model was optimized for different ecosystems by addressing three types of parameters to improve the regional simulation accuracy. Based on flux observation data， remote sensing imagery， and related spatial data， an ecosystem respiration equation was then reconstructed using flux station observations to calculate the accurate Net Ecosystem Productivity （NEP） of Xuzhou. A spatio-temporal analysis of the GPP/NEP relationship in Xuzhou was conducted. The results showed that the optimized model significantly improved simulation accuracy， with an average correlation of R=0.85 and a Root Mean Square Error（RMSE） of 2.03 gC·m⁻²·d⁻¹ between observed and simulated values. The average NEP in the study area over 14 years was 1.56 Mt C·a⁻¹， indicating that the region has maintained its carbon sink function. The spatial distribution of the carbon sink shows stronger sequestration in the densely vegetated areas in the north and northwest， with weaker sequestration in the urbanized areas in the central and southeastern regions. This study gives important insights for accomplishing the carbon goals for carbon sequestration capacity and regional ecological management of Xuzhou city.
Semantic segmentation of LiDAR 3D point clouds is fundamental for complex 3D analysis and applications. High-precision point cloud semantic segmentation is important for applications such as robotics， autonomous driving， smart cities， and augmented reality. Recently， deep learning-based methods have been proposed for 3D point semantic segmentation task. However， due to the complexity of street scene， the existing methods still have the problems of limited receptive fields which severely limit the semantic segmentation performance. To solve this problem， this paper proposes a deeply supervised multi-scale self-attention network suitable for street scene semantic segmentation in 3D point cloud. The proposed network adopts multiple sub-encoders with different receptive fields to extract features in different scales. To further enlarge receptive fields without introducing high computation cost，Dilated K-Nearest Neighbors（DKNN） with different dilation rates are utilized in each sub-encoder. Self-attention convolution is used as the basic unit and deep supervision is introduced in the network to accelerate convergence. Experimental results show that the proposed network achieves the best performances with 97.2% OA and 82.2% mIoU on the challenging Toronto 3D dataset， and 95.1% OA and 70.3% mIoU on CSPC dataset.
Cyanobacterial blooms pose serious threats to aquatic ecosystems and human health. Operational satellite-based monitoring of cyanobacterial blooms， with rapid response to management needs， is of significant scientific and practical value. Current research mainly focuses on automated extraction algorithms and spatiotemporal variation analysis， while high-precision， standardized， and operational methods for rapid spatiotemporal response are limited. Consequently， monitoring results often cannot effectively support national and local water environment management. To address this， we developed a comprehensive methodology for satellite-based operational monitoring of cyanobacterial blooms， adhering to principles of scientific rigor， operability， and comparability. The methodology includes satellite imagery selection， monitoring frequency and response times for routine and emergency monitoring， bloom distribution extraction， and quality control of results. Based on this， a product system for operational rapid spatiotemporal monitoring was established. This approach improves the accuracy， consistency， and timeliness of cyanobacterial bloom monitoring and provides critical support for water environment management and decision-making at national and local levels.
