Fire threatens the safety of human life and property and causes great damage to ecosystems. The study of remote sensing response characteristics of burned area is important for the accurate extraction of area， quantitative assessment of fire damage and vegetation restoration. Based on Sentinel-1 SAR remote sensing images， the characteristics of unburned forest， burned area， buildings， water bodies were analyzed in six fire cases. The time series of burned area from one year before the fire to two years after it was analyzed. The results shows that the cross-polarization ratio and the backward scattering intensity of VH polarization are lower in the burned area compared to the unfired area， the backward scattering intensity of VH and VV polarization of buildings are much higher than those of other features， and the backward scattering of water bodies in both polarizations is very low， while the cross-polarization ratio is higher. From time series perspective， the backward scattering intensity for VV polarization shows obvious seasonal variations. In most cases， the backward scattering intensity for VV polarization is significantly higher within one month after the fire， and the cross-polarization ratio rapidly decreases. The time series variation of Normalized Burned Ratio index （NBR） calculated from Sentinel-2 MSI follows a consistent pattern with SAR images， showing obvious seasonal changes. It rapidly decreases within half a month after the fire and gradually recovers.
Pine wilt disease is a devastating pine disease， which seriously threatens forest ecological security. Timely and reliable acquisition of the extent and severity of pine wilt disease is very important for forest management and disease prevention and control. However， pine wilt disease spreads rapidly and is difficult to control， and the traditional manual survey methods can hardly meet the demand. Unmanned Aerial Vehicle （UAV） remote sensing can quickly and accurately obtain the extent and severity of forest diseases， and provide reliable information support for forest pest control and management. In this study， UAV was used to acquire high-resolution Red-Green-Blue （RGB） visible light images. Firstly， object-oriented multi-scale segmentation algorithm was used to extract the crown of a single tree， and Vegetation Index （VIs） and texture （GLCM） features were calculated. Then the feature selection algorithm was used to optimize the feature set， and Random Forest （RF） classification and Support Vector Machine （SVM） classification algorithm were used to construct the pine wilt disease classification model based on different feature sets. Through the ablation experiment， the optimal classification model was selected and the object-oriented method was used to monitor the disease degree and spatial distribution of pine wilt disease. The results show that the vegetation index and texture features of pine canopy with different disease degrees are different on the object-oriented single tree crown scale， and the accuracy of classification results using vegetation index was better than texture characteristics （VIs RF：OA=76.52%，Marco-F1=0.77；SVM：OA=79.68%，Marco-F1=0.79）. Compared with a single feature set， the combination of vegetation index and texture features can significantly improve the classification accuracy （VIs&GLCM RF：OA=79.47%，Marco-F1=0.80；SVM：OA=85.45%，Marco-F1=0.85）， indicating that multi-feature combination can effectively improve the pine wilt disease classification. The SVM model outperforms the RF model for classification， both for single feature set modeling and combined feature set modeling. This study provides timely and reliable information to support a comprehensive grasp of the extent and severity of pine wilt disease， and helps to promote the construction of a major forestry pest control system and maintain ecological security.
Understanding the dynamics of Leaf Area Index （LAI） in tropical evergreen forests is crucial for assessing ecosystem health and carbon dynamics. In this paper， a method is proposed for decomposing LAI into leaf age cohorts in tropical evergreen broadleaf forests across Amazon Basin. The method simplifies the canopy into three major leaf age stages （i.e.， young， mature， and old leaves）. The method integrates leaf-level photosynthetic biochemistry models with remote sensing climate data to simulate carbon assimilation across these leaf age stages. Then， utilizing a novel neighbor-based approach and the linear least squares solver with bounds or linear constraints （Lsqlin） to derive the values of three LAI cohorts. Validation against ground-based phenology camera data shows good agreement in seasonal dynamics of LAI cohorts （LAIyoung： R2 = 0.32； LAImature： R2 = 0.61 and LAIold： R2 = 0.49）， indicating the method's ability to capture seasonal variations accurately. Spatial patterns of LAI cohorts closely correspond to climatic phenology variables across the Amazon Basin. This approach enhances our understanding of LAI dynamics in tropical forests， providing valuable insights for ecosystem management and carbon cycle modeling in the Amazon Basin.
In order to realize more accurate mangrove extraction and monitoring， four mangrove plantation areas， including the Aojiang River coast in Wenzhou City， were taken as the study area， and the distribution of Mangrove forests was extracted and accuracy verified based on the DeepLabV3+ semantic segmentation model using Sentinel-2 remotely sensed imagery data， which was applied to analyze the spatial change of mangrove forests in the period of 2019~2023. The results show that： ① the Mangrove information extraction model constructed by DeepLabV3+ network can better distinguish Mangrove and Non-Mangrove areas， with fewer mis- and omissions； ②the semantic segmentation algorithms are significantly better than traditional machine learning methods， with the DeepLabV3+ method having the highest accuracy， with an precision of 84.89% and a Kappa coefficient of 0.82； ③The growth of Mangrove forests is greatly affected by the geographical location and growth environment， and Mangrove forests in the intertidal zone of the coast or at the mouth of the sea are more susceptible to the influence of typhoons， tides， etc.， and the encroachment of mangrove forests' growth space by exotic species， such as spartina alterniflora， etc.， are all the key factors that cause the low survival rate of Mangrove seedlings and the slow growth rate. Therefore， the semantic segmentation model based on DeepLabV3+ can better recognize and extract the Mangrove forests and provide data base support for the monitoring and assessment of Mangrove forests in Wenzhou City.
Forest age significantly affects trends of forest carbon sink. The spatial data of forest age are required to reduce uncertainties in regional and global forest carbon sinks. Forest age and canopy height are closely linked. In recent years， more and more high-resolution forest canopy height data derived from remote sensing available， making it possible to map forest age at a high-resolution. However， the feasibility of high-resolution mapping of temperate forest age from remotely sensed forest height. Therefore， studying the estimation and mapping of temperate forest age based on forest height remote sensing data is of great significance for improving the accuracy of regional carbon sink dynamics monitoring， optimizing forest management strategies， and deepening the understanding of carbon sequestration mechanisms in temperate forest ecosystems. With Heilongjiang Province as the study area， the optimal growth equations describing the changes of canopy height with forest age were determined for different forest types， including deciduous broadleaf， evergreen needleleaf， deciduous needleleaf， and mixed forests and the plot data were corrected for time， based on measurements at 1821 plots. 70% of plot data were randomly selected for model training and remaining 30% of data were used for model validation. With forest height derived from LiDAR data and environmental factors （including the length of the growing season， the highest monthly average temperature， and slope） as the independent variables， models estimating forest model were constructed using the Random Forest （RF）， the Support Vector Machine （SVM）， and the LightGBM methods， respectively. The best model was used to map forest age in 2020 at a spatial resolution of 30 and the characteristics of forest age in the study area was analyzed. The results showed that for the training and validation samples， the RF model achieved the highest R2 （0.77） and lowest root mean square error （RMSE=10.20）， followed by LightGBM model. The SVM model achieved the lowest R2 （0.63） and highest RMSE （11.85）. There were obvious spatial variations in the forest age estimated using the RF model. Forest age were significantly higher in Daxinganling district and Yichun city than in other regions， and were low in Heihe city. Deciduous coniferous forests had the highest average age， followed by evergreen coniferous forests and mixed forests. Deciduous broadleaf forests had the lowest average age. Forest ages in the study area averaged 73 years， with 75% of the forests aged between 40 and 100 years. Forests older than 100 years accounted for 17%， and 8% of the forests were younger than 40 years. The study shows that the age of temperate forests in China can be estimated using a machine learning method by combining remotely sensed forest height with environmental factors， which was valuable for mapping regional and global forest age at a high spatial resolution.
In order to meet the demand for accurate monitoring of forest resources， this paper explores the potential of backpack Light Detection and Ranging （LiDAR） on extracting forest structure parameters for the practical applications. Taking Jiande Forest Farm in Zhejiang Province as the study area， based on backpack LiDAR data collected from eight sample plots， an improved K-means hierarchical clustering algorithm is proposed for individual tree segmentation. Then， six individual tree structural parameters， including diameter at breast height， tree height， crown diameter， crown area， crown volume and gap fraction， as well as 56 cloud point layer height variables were calculated based on the segmented individual tree point cloud data. After that， the random forest method is applied to estimate the volume of individual trees and sample plots. The results showed that， the average comprehensive segmentation accuracy F of the improved K-means hierarchical clustering algorithm was 0.87， and the extraction accuracy of single tree diameter at breast height was 91.26%， and the tree height accuracy was 85.77%. The individual tree volume estimation model using only six tree structural parameters obtained an accuracy of the coefficient of determination（R²） of 0.89， and the Root-Mean-Square Error （RMSE） was 0.053 m3. After using the Pearson correlation coefficient and the importance of random forest features to select the optimal features from the individual tree structure parameters and layer height parameters， the outperformed model was obtained with an estimation accuracy of R² was 0.93， RMSE was 0.041 m3， and the overall plots’ accuracy reached 94.20%. This study indicated that the proposed K-means hierarchical clustering algorithm can effectively segment individual tree point clouds， and the random forest method can estimate individual tree volume and sample plots volume well， which can provide an important reference for backpack LiDAR in extracting forest resource parameters.
LiDAR point cloud data serves as a crucial data source for forest resource inventory. However， the registration of multi-view terrestrial LiDAR point cloud data in forest scenarios is often plagued by inefficiency. Addressing the limitations of current research， this study proposes a target-free and rapid point cloud registration method based on point cloud normal features as the feature descriptor. Firstly， the original point cloud is subjected to denoising and voxelization processing. Then， point cloud normal vectors are computed， and feature matching is performed. Finally， precise registration is achieved using the nearest neighbor iterative algorithm. The proposed method is tested in field plots with different vegetation characteristics thus registration difficulty levels. Experimental results show that the optimal voxelization sampling interval is between 30 cm and 50 cm. Compared to the reference results derived from manual registration， the average horizontal translation error and average vertical translation error are 3.13 cm and 0.86 cm， respectively， while the average rotation error is 1.39 '. The average processing time is 5.2 s， and the average point-wise error is 6.5 cm. This method successfully improve the efficiency and accuracy of automated point cloud registration in complex forested environment.
To address the limitations of traditional single-tree Diameter at Breast Height （DBH） measurement methods， including inefficiency， restricted accuracy， and systematic underestimation of parameters due to point cloud data incompleteness， this study proposes an optimized approach integrating hierarchical slicing strategies with multi-model collaboration to enhance the precision and applicability of 3D laser point cloud technology in DBH extraction. A multi-thickness hierarchical slicing framework was designed for the critical DBH measurement interval （1.2~1.4 m）， coupled with a comparative analysis system incorporating least squares， Hough transform， and RANSAC circle-fitting models， validated through field-measured data. Experimental results demonstrated that the hierarchical strategy reduced the Mean Absolute Error （MAE） and Root Mean Square Error （RMSE） of full-scene DBH extraction by 18.7% and 22.3%， respectively. In complex scenarios， the RANSAC model achieved the highest coefficient of determination （R2） of 0.980 5， representing a 14.6% improvement over conventional methods， while the Hough transform and least squares methods attained R2 values of 0.779 1 and 0.969 1， respectively. The study confirms that hierarchical slicing mitigates fitting inaccuracies caused by data incompleteness through optimized point cloud density distribution， with the RANSAC model exhibiting superior robustness for irregular point cloud fitting. This methodology provides a reliable technical solution for dynamic forest resource monitoring and carbon sequestration quantification， advancing the application of 3D laser scanning in smart forestry management.
Forest canopy closure is a crucial parameter in forest resource inventory that plays a significant role in evaluating and monitoring the stability of forest ecosystems. With the continuous development of remote sensing technology， the estimation of large-scale forest canopy closure using multi-source remote sensing data has become a hot research topic. In this study， a regression model was constructed using machine learning algorithms based on laser point cloud data and multi-source optical remote sensing data to estimate forest canopy closure in large forested areas. Firstly， the dependent variable of the regression model， which is the true values of forest canopy closure， was calculated from the Airborne Laser Scanning （ALS） point cloud data. Secondly， 18 independent variables， such as vegetation indices and texture， were extracted from Sentinel-2 MSI， Landsat-8 OLI， and Sentinel-1 SAR images. Then， taking 14 forest plots in Guangxi as an example， the impacts of different independent variable combinations on forest canopy closure inversion were experimentally analyzed using two machine learning models， Random Forest Regression （RFR） and Support Vector Regression （SVR）. Finally， the best variable combination and machine learning method were selected to map the forest canopy closure in Guangxi. The experimental results showed that RFR performed better than SVR， and the S2+S1 combination had the highest accuracy， with a correlation coefficient R2 of 0.703， Root Mean Square Error （RMSE） of 0.19， and Mean Absolute Error （MAE） of 0.13. Additionally， polarization features can significantly improve the inversion accuracy of forest canopy closure.
Considering that the Geosynchronous Interferometric Infrared Sounder （FY-4A/GIIRS） has high spectral and radiometric calibration accuracy and operates on Geostationary orbit （GEO）， GIIRS can provide higher temporal frequency radiometric comparisons of other broad-band infrared remote sensors including those in Low-Earth orbit （LEO） and GEO to monitor the intraday variation of their radiometric performance. Therefore， based on the inter-calibration baseline algorithm issued by the Global Space-based Inter-Calibration System （GSICS）， a method for radiometric reference transfer with GIIRS as the reference sensor is established. This method defines inter-calibration conditions and thresholds suitable for different platform instruments between LEO-GEO and GEO-GEO， based on the characteristics of high observation frequency but low spatial resolution of GEO sensor， which can support inter-calibration with higher time frequency and larger regional scope. It is shown through the radiometric reference transfer of the Advanced Geosynchronous Radiation Image （FY-4A/AGRI） that， it is feasible to use GIIRS for radiometric reference transfer， and the collocated results of long-wave infrared band （10.3~11.3 μm） have a good fitting relationship in most times. The brightness （BT） deviation of AGRI relative to GIIRS is increased because of too few sampling points. After removing the result of too few points （less than 1 000）， the average absolute BT deviation is 1.2 K@290 K， which is 0.7 K lower. The BT deviation fluctuates slightly at odd times and greatly at even times. The main reason is that the GIIRS observation area at odd times is in low latitudes （10°~30°N）， while it is in middle and high latitudes （30°~60°N） at even times. As the latitude rises， the spatial collocation error increases and relatively uniform collocated pixels decreases significantly with the footprint distortion increases. In order to reduce the error， when calculating the radiance of the target area， it should be variably weighted （instead of uniformly weighted） according to the relative positions of the pixels， so as to improve the collocation accuracy when GIIRS is used as a reference remote sensor.
Accurate estimation of leaf or canopy chlorophyll content is essential for monitoring crop growth. Crop chlorophyll monitoring using remote sensing is a non-destructive， large-area，real-time monitoring method， which requires reliable inversion models and satellite data. Focusing on summer maize， parameter settings of PROSAIL model were determined through local and global sensitivity analysis. In combination with measured ground data and related literature， canopy reflectance of summer maize based on PROSAIL model was simulated. Then， based on spectral response function of Sentinel-2A image， equivalent reflectance data of Sentinel-2A image were obtained. Typical hyperspectral vegetation indices and vegetation indices of improved band combination mode based on Sentinel-2A image data were calculated and analyzed to determine the best estimation model of Leaf Chlorophyll Content （LCC） and Canopy Chlorophyll Content （CCC）. Based on PROSAIL simulation data， Sentinel-2A image data and ground measurement data， the modeling and validation analysis of summer maize LCC and CCC were carried out. The results showed that the R2 of vegetation indices inversion based on PROSAIL model and Sentinel-2A image were 0.61 and 0.65， RMSE were 7.54 μg/cm² and 8.46 μg/cm²， respectively. The inversion accuracy of LCC using vegetation indices based on PROSAIL model and Sentinel-2A image was consistent， and the inversion accuracy met the requirements of summer maize growth monitoring. The R2 of CCC retrieved by the two methods were 0.75 and 0.77， RMSE were 1.03 g/m2 and 0.02 g/m2， respectively. This study provides an effective method for retrieving chlorophyll content in the region where there are few ground measured data， which is helpful for the growth and pest control monitoring of summer maize.
Landslide susceptibility assessment is a crucial tool for proactively preventing and controlling landslide disasters and avoiding casualties and property damage. This letter proposes a dynamic method for calculating landslide susceptibility， which aims to quantify the influence of the rainfall as the main inducing factor， on the landslide deformation using the time–frequency analysis method， allowing the calculation of landslide susceptibility on a time scale. Firstly， the rainfall and the surface deformation data from Interferometric Synthetic Aperture Radar （InSAR） are performed by the wavelet analysis to quantify the response of surface deformation to rainfall. Secondly， the rainfall after calculating the quantitative relationship and other landslide elements are fed into the designed Random Forest model （RF）. Finally， the landslide sensitivity is compared across time to assess the timeliness of the model. The results demonstrate that the landslide sensitivity model can effectively predict the landslide risk changes and high predictive accuracy. The AUC is 0.962， with accuracy and precision rates of 0.916 and 0.941. Furthermore， the model presented demonstrates significant discriminative power in time series analysis. and high-risk areas primarily concentrated around the known landslide hazards， suggesting that the predictions of the model coincide with the distribution of actual landslide hazards.
The Liaohe Estuary coastal wetland is the northernmost estuary wetland in China， which is an ideal breeding and migration station for many kinds of waterfowl. In recent years， several ecological restoration projects have been carried out to improve the habitat quality in this region. Accurately mapping the landcover types using remote sensing is very important for efficiently evaluating wetland habitat quality and restoration effectiveness. However， most of the classification methods in the Liaohe Estuary were object-oriented， and the mapping results were not fine enough and needed to be updated in years. The applicability of pixel-level method and dense time-series information in this region needed to be further evaluated. This paper relied on Google Earth Engine （GEE） platform， utilized Sentinel-2， Sentinel-1， and topographic multi-source data to extract the features including spectral indices， texture， topography， backscattering， and phenology from the dense time-series vegetation indices. Multi-year sample datasets were generated by field sampling and sample migration， and the pixel-level fine classification mapping from 2018 to 2022 was carried out based on the random forest model. The effects of different features on the classification accuracy were also evaluated. The classification method combining GEE and dense time-series information got an overall classification accuracy of 95.77%. Adding phenology features improved the accuracy most obviously， especially for the mixing between suaeda salsa and reeds， rice， or aquaculture ponds. Adding texture and backscattering features significantly improved the accuracies of aquaculture ponds and construction land. In the last five years， aquaculture ponds decreased while the mudflat and suaeda salsa expanded， which indicated the effects of ecological restoration project. The results provide data and technology supports for analyzing the spatial-temporal changes and driving mechanism of coastal wetland， which is of great significance for strengthening the protection and restoration of wetland ecosystems.
The oblique photogrammetry-based 3D modeling pipeline cannot reconstruct perfect traffic sign models， although it has been widely applied in large-scale 3D urban modeling. Therefore， a deep learning-based traffic sign embedded modeling method is proposed to fix the traffic sign 3D modeling issue using oblique photogrammetry. First， a deep neural network is used to detect the traffic signs in oblique aerial images. Second， taking advantage of the detection results， a template matching method is applied to generate the traffic sign 3D models with completed structure and perfect texture. Third， under the geometric constraints placed by the detected bounding boxes， the Scale-Invariant Feature Transform（SIFT） is used to extract the corresponding points on the traffic signs. Additionally， based on stereovision， triangulation is applied to obtain the 3D point of a single traffic sign in the city scene. Last， least-squares fitting is used to the refined point cloud to fit a plane for orientation prediction. The road signs with computer-aided design models are embedded in the 3D urban scene. The experimental results show that the proposed method achieves a high mAP in traffic sign detection and produces visually plausible embedded results， demonstrating its effectiveness for traffic sign modeling in oblique photogrammetry-based 3D scene reconstruction.
GRACE satellite has opened a new era of quantitative retrieval of groundwater change by remote sensing， but it has the problem of low spatial resolution. High-resolution groundwater observations will significantly improve the accuracy of local-scale hydrological process understanding， thereby offering essential data support for the development of scientifically based groundwater management policies. After sorting out the characteristic factors such as precipitation， air temperature， evapotranspiration， surface temperature， normalized vegetation index and soil water in the Yellow River Basin， partial least squares regression method was used to screen the characteristic factors respectively from January to December， and the optimal monthly characteristic factor subset was constructed. Then， the random forest algorithm was used to downscale the groundwater data of the Yellow River Basin from 0.25°× 0.25° to 1 km×1 km， and compared and verified with the measured groundwater level data. The results show that： （1） Except evapotranspiration and surface temperature， the importance of other factors changes with the change of month； （2） In the time series， the correlation coefficient and Nash coefficient of groundwater data before and after downscaling are as high as 0.95， and the root-mean-square error is 3.17 mm； （3） Spatially， compared with before downscaling， the correlation coefficient between the change data of groundwater reserves and the measured groundwater level after downscaling increased by 47.67%. The research results can meet the demand for high-resolution groundwater data in practical applications， and provide reference for the feature factor screening of groundwater downscaling research.
Savannas， characterized by their low vegetation density and substantial total aboveground biomass， represent a critical region for global carbon cycle. Nevertheless， significant spatial heterogeneity exists within these ecosystems， leading to considerable uncertainty in remote sensing biomass estimations. The Global Ecosystem Dynamics Investigation （GEDI） provides high-quality estimates of above-ground biomass within its observed footprint by leveraging three-dimensional surface vegetation information from LiDAR. However， it lacks spatial continuous above-ground biomass data. Sentinel-2， PALSAR-2 and tree cover data were used to extract 28 features， and a random forest model was established with GEDI footprint level above-ground biomass data to build a high-resolution above-ground biomass estimation method for African savanna. The results show that the algorithm can generate spatially continuous above-ground biomass data in study area， and effectively extract tree information in non-forested areas that are often ignored in previous studies. The mean absolute error and root-mean-square error of the model are 15.798 Mg/ha and 24.626 Mg/ha， respectively. The accuracy remains consistent when using optical images from different seasons. When modeling with optical data acquired during rainy season， spectral bands such as red， red edge， and short-wave infrared， along with their relative spectral indices， play crucial roles. In contrast， when using dry season optical data， tree cover and InSAR become significantly more important. When conducting large-scale biomass estimation of African savannas， the use of multiple data sources can help to obtain better estimation accuracy. This study provides a method for low-cost monitoring of aboveground biomass in Savannas in the future， and contributes to the in-depth study of vegetation carbon cycle in this region.
The agricultural and pastoral land on the Qinghai-Tibet Plateau forms the basis for safeguarding natural grasslands and maintaining the ecological security barrier in the region. There is an urgent need for precise characterization of the spatial distribution pattern of this land. Satellite remote sensing technology is extensively used to rapidly and accurately generate spatial distribution maps of land cover. This approach offers a technical means for identifying agricultural and pastoral land. In this study， the Google Earth Engine （GEE） cloud platform is utilized， incorporating phenological knowledge and machine learning algorithms. Ground-truth data， Sentinel-1 Synthetic Aperture Radar （SAR） imagery， and Sentinel-2 optical remote sensing imagery serve as data sources. The study identifies the distribution of typical mixed agricultural and pastoral land with barley and oilseed rape cultivation on the Qinghai-Tibet Plateau by analyzing radar polarization features， optical vegetation index characteristics， and topographical features. The results of the study show that the total planting area of barley and oilseed rape in Shigatse City during 2019~2023 shows a trend of steady growth， the planting structure is relatively stable， and the distribution pattern shows obvious characteristics of more in the east and less in the west， and the overall distribution is more dispersed. Combining Sentinel-2 optical remote sensing data with Sentinel-1 SAR radar data in the classification process significantly improved the overall accuracy， Kappa coefficient and F1 score of the classification compared to using only a single data source feature. Further combination of topographic features resulted in another improvement in accuracy and a closer match between the remotely sensed estimated planted area and the actual area in the statistical bulletin. In view of this， by integrating the vegetation index， topography and backscattering features， the study achieved accurate identification of barley and oilseed rape plantations， with overall classification accuracies exceeding 92% during the period， with the lowest Kappa coefficients of 0.841 and F1 scores higher than 0.917.This study establishes a crucial methodological foundation for mapping the distribution of artificial grassland cultivation on the Qinghai-Tibet Plateau and scientifically formulating policies for grassland livestock development and ecological conservation.
The identification of urban functional zones can assist in the decision-making process in urban construction. This paper proposes a multi-source scene feature fusion method utilizing the Transformer model for urban functional zone identification. Firstly， the Traffic Analysis Zone （TAZ） is constructed based on the road network. The graph structure of POI （Point of Interest） data is created using the Delaunay Triangulation （DT）. Additionally， remote sensing data is utilized to obtain the corresponding image objects for each TAZ. Subsequently， the POI graph structure is processed using a Graph Convolution Network （GCN） to extract social scene features. Meanwhile， the natural scene features of remote sensing data are obtained through encoding with ResNet-50. Finally， the multi-head attention mechanism of Transformer decoder is utilized to fuse multi-dimensional feature vectors， facilitating accurate identification of urban functional zone with SoftMax. Taking the main urban area of Shenyang as an example， multi-source data such as OSM （Open Street Map）， POI and remote sensing data in 2021 are used as experimental data. The results indicate that the overall accuracy and Kappa coefficient of this method are 82.2% and 70% respectively. Furthermore， the Kappa coefficient is at least 18% higher than that of the single data method and at least 9% higher than that of other fusion methods. This study innovatively employs the Transformer model to integrate social and natural scene features， effectively addresses the challenge of combining diverse features from multiple sources into an integrated representation， and provides a new technical approach for urban functional zones identification.
The interaction between human activities and land cover has important impacts on ecosystems. A study was conducted on the land cover of Shanxi Province to estimate the Human Activity Intensity Index of the Land Surface （HAILS） for the years 2015 and 2020. Spatial analysis was employed to explore the spatial and temporal distribution characteristics of the HAILS in terms of spatial pattern， slope and water flow paths， and correlation indices were used to discuss the validity of the HAILS and explore the influencing factors of its changes. The results indicate that the areas with higher HAILS from 2015 to 2020 were predominantly distributed in the basin and along the Fen River， and the percentage of areas with HAILS values greater than 20% increased in all slope bands. The water flow path results show that the HAILS values of the Fen River and other rivers decreased significantly after 2 km， and the HAILS values of the Fen River were highest both before 20 km and after 28 km.The HAILS was significantly correlated with the actual water consumption and the first， second and third GDP， but weakly correlated with the population density， the GDP and the night light data. Consequently， economic development， industrial structure and ecological protection policy are identified as the primary factors influencing the change of human activity intensity.
Based on Sentinel-2 images in the core area of four quaternal phases Yancheng Wetland Rare Birds National Nature Reserve in 2020， spectral features， texture features， red edge index， vegetation and water index characteristics were extracted， and 14 sets of information extraction schemes， including single-season facies， multi-seasonal facies， and preferred feature combinations based on vegetation growth rules， were designed to compare the classification effects of K nearest neighbor and random forest two machine learning methods. The results show that the single-season phase classification accuracy and stability of random forest are higher than those of K nearest neighbor. The addition of texture features improves the accuracy of vegetation classification. The overall classification accuracy of the preferred feature combination in the vegetation growth period reached 98.93%， Kappa coefficient was 0.986， and the overall accuracy of vegetation dormancy period was 97.97%， and the Kappa coefficient was 0.978， which verified the effective correlation between vegetation growth law and information extraction results.The purpose of this paper is to use the optimization scheme to use the optimization scheme to improve the misseparation caused by mixed cells when extracting vegetation information by remote sensing technology， and the redundant information generated by the combination of multiple features of dimensionality reduction， which can provide certain reference value and technical assistance for the monitoring and protection of Yancheng coastal wetlands.
The continuous changes in land use patterns have profound impacts on the ecological environment and socio-economic development of the Yellow River Basin. Understanding the dynamic changes in land use in the middle reaches of the Yellow River Basin can provide a scientific basis for achieving sustainable development in the Yellow River Basin. This study is based on four periods of Landsat remote sensing data from 1995， 2005， 2015， and 2023 in the middle reaches of the Yellow River Basin. Obtain the spatial distribution of land use in the demonstration area through Support Vector Machine （SVM） and maximum likelihood method classification， and analyze the characteristics of land use change in the study area from 1995 to 2023 using quantitative indicators such as land use change and transition matrix. The Markov model was applied to predict the land use changes in 2025 and 2030. By establishing the FLUS-Markov model， the land use changes in the study area in 2025 and 2030 were predicted. For the six types of land use changes （forest， grassland， wetland， cultivated land， construction land， and unused land）， the results show that： （1） Over the past nearly 30 years， cultivated land and forest land have decreased by 8 600 km² and 6 400 km² respectively， while the area of construction land has significantly increased by 7 500 km²； （2） The transfer directions of land use types are diverse， mainly from cultivated land to construction land and vegetation， and the landscape of each land type tends to be balanced in spatial distribution， with coordinated urban development； （3） Between 1995 and 2023， the number of forests， grasslands， and construction land showed an upward trend， while the number of wetlands， cultivated land， and unused land showed a downward trend； （4） In the next 10 years， there will be significant changes in land use in the middle reaches of the Yellow River， as humans require more land for construction.
