Deep learning has significantly advanced remote sensing image processing technology， demonstrating notable improvements in both accuracy and speed. However， deep learning models typically require large amounts of manually labeled training samples in practical applications， and their generalization performance is relatively weak. In recent years， the development of visual foundation models and large language models has introduced a new paradigm for research on large models in remote sensing image processing. Remote sensing large models， also known as remote sensing foundation models， have garnered attention for their outstanding transfer performance in downstream tasks. These models are first pretrained on large datasets unrelated to specific tasks and are then fine-tuned to adapt to various downstream applications. Foundation models have already been widely applied in language， vision， and other fields， and their potential in the field of remote sensing is increasingly gaining attention from the academic community. However， there is still a lack of comprehensive surveys and performance comparisons of these models in remote sensing tasks. Due to the inherent differences between natural images and remote sensing images， these differences limit the direct application of foundation models. Against this backdrop， this paper provides a comprehensive review of common foundation models and large models specifically designed for the field of remote sensing from multiple perspectives. It outlines the latest advancements， highlights the challenges faced， and explores potential future directions for development.

With the acceleration of urbanization and industrialization in China， the urban thermal environment has undergone tremendous changes， and the urban heat island effect is gradually strengthening， which has adverse effects on the urban ecological environment and climate. This paper calculated Surface Urban Heat Island Intensity （SUHII） by grids using MODIS Land Surface Temperature （LST） of Taiyuan main urban area from 2003 to 2021， analyzed the spatiotemporal distribution changes of Surface Urban Heat Island （SUHI） and their relationship with urban expansion using spatial statistical methods， and then explored the influencing factors on SUHI based on random forest model. The results showed that： （1） In the past 20 years， the heat island effect in Taiyuan City has shown a growing trend with urban development， with significant seasonal differences， with the strongest in summer and the weakest in winter. （2） The urban heat island effect in the main urban area of Taiyuan City has significant positive spatial autocorrelation， and the highly agglomeration area is significantly expanded. （3） The spatial expansion direction of urban heat islands is basically the same as the direction of urban expansion.（4） Human factors are the main factors affecting the urban heat island effect， with GDP and PM_2.5 having the greatest impact. This study can provide a methodological reference for for the quantitative evaluation of urban effects in areas with significant terrain fluctuations， and provide a understanding of the thermal environment in Taiyuan City and a scientific reference for formulating urban planning strategies.

The semantic segmentation of remote sensing images by using machine learning and deep learning is an important means in the field of intelligent interpretation of remote sensing. Optical and SAR images can reflect different land surface characteristics and provide additional information for land use/cover. Effectively integrating and improving the recognition of land use/cover types is a major difficulty in remote sensing semantic segmentation. To solve these problems， a land use/cover segmentation method based on multi-source data fusion and channel correlation perception was proposed. Firstly， using Gaofen-6 optical image and Gaofen-3 radar image to produce high-resolution image sample data； Then， channel attention mechanism and spatial multi-scale correlation are introduced into deeplabv3p framework using coding-decoding architecture， and CCAMNet（Cross-channel sensing Module network） model based on full convolutional network structure is established. Finally， a comparative experiment is conducted with UNet， Deeplabv3p， SA-Gate， v-fusenet， MCANet and other semantic segmentation models on the research region data set. Moreover， it is compared with Deeplabv3p， PSCNN， MRSDC， v-fusenet， MBFNet， MCANet and other models on WHU-OPT-SAR public data set. The results show that compared with other models in the study area， the overall accuracy of the proposed model is 79.68%， which is better than other models. On the open data set， the overall accuracy is 81.8%， which is better than other models in classification accuracy and can significantly improve the accuracy of semantic segmentation， and verify the feasibility of the model in the application of semantic segmentation in remote sensing land use classification. The research results filled the gap of the lack of high-resolution land use classification dataset in the study area， and verified the feasibility of optical and SAR fusion in significantly improving classification accuracy.

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.

Polar sea ice， with its important impact on global climate change， makes accurate acquisition of multi-element information of sea ice the core task of polar observation. Satellite is the main technical means of polar sea ice monitoring， which has been widely used to observe polar sea ice at the domestic and foreign. To clarify the current status of satellite remote sensing of polar sea ice at home and abroad， which is an important guideline for the development of new remote sensing sensors for sea ice in polar regions in the future. In this paper， the domestic and foreign satellites with polar sea ice information acquisition capability that are currently in orbit are reviewed in detail. On this basis， the main application progress in polar sea ice observation based on satellite data is summarized. Finally， it points out the shortcomings of the existing global earth observation system of polar sea ice， and puts forward suggestions for the development of China's subsequent polar sea ice observation.

Forest age is an important characteristic parameter of forest ecosystem， and it is of great significance to accurately estimate carbon storage and carbon sink of forest ecosystem. However， at present， there are few reviews on the age datasets and its estimation algorithm， so this paper systematically summarizes and analyzes the existing forest age datasets， which were divided into global and regional forest age datasets according to the spatial coverage of the data. Then， we analyzed the algorithms of forest age datasets and their advantages and disadvantages. Research has shown that （1） The forest age algorithms mainly include the downscaling statistical method， relationship equations between forest age and forest structure parameters， forest disturbance detection algorithms and machine learning algorithms （such as random forest algorithm）. （2） The advantage of the downscaling algorithm is simple and easy to use， but its main disadvantage is that the saturation of normalized differential vegetation index will underestimate the age of the old-growth forest. The advantage of the relationship model between forest age and forest structure parameters is that the high-precision remote sensing data of tree height or biomass data can reflect the spatial heterogeneity characteristics of forest age， and the forest growth model has a theoretical basis. But the disadvantage is that the results are restricted by the accuracy of forest or tree species distribution map， and the influence factors of forest growth are not considered comprehensively. The advantage of forest disturbance monitoring algorithm is that mature algorithms can be used to detect forest disturbance and infer the change of forest age， but the disadvantage is that the forest age must be combined with other methods to obtain the age of old forest. The advantage of random forest algorithm model is that the model is easy to build， does not need to set specific statistical assumptions and model forms， and doesn’t rely on forest type map or tree species distribution map. The disadvantage of this method is that it is restricted by the numbers and the spatial distribution representation of the model training samples. （3） Forest inventory data and remote sensing data are important data for forest age estimation. Forest age has important application prospects in ecological model driving， forest management and carbon neutrality. In the future， the research of forest age should strengthen the ground observation of forest age， combine the advantages of remote sensing data， and use a variety of machine learning algorithms to develop high spatial and temporal resolution forest age data.

Black and odorous water occur frequently in rural China， and research into monitoring them by employing remote sensing technology has only recently begun， with many technical issues to be resolved.The samples were collected in rural areas of Jilin， Yunnan， and Guangxi provinces for this study， and 75 water samples from black and odorous water and 85 water samples from normal water were collected between 2021 and 2022， and their water quality parameters as well as optical properties were analyzed separately. We analyzed the image spectral characteristics of black and odorous water and normal water using GF-2 images， and observed that the reflectance of rural black and odorous water has an increasing trend in the red and near-infrared bands， whereas the reflectance of red and green bands was very low and the difference was small. Based on the two typical spectral characteristics of black and odorous water， the MBOI （Multi-spectral black and odorous water index） was developed， with a high identification accuracy. The following are the main research findings： （1） Black and odorous water have a higher concentration of total suspended particulate matter than normal water， and the concentration of organic carbon in black and odorous water bodies is 1.82 times higher compared to normal water. （2） At 440 nm， the absorption coefficients of all the materials， including algal particulate matter， non-pigmented particulate matter， and colored dissolved organic matter of black and odorous water were greater than those of normal water. （3） The data modeling and model verification are carried out by using the spectral reflectance data of the image after Rayleigh correction. When the MBOI value is between 0 and 0.18， it is determined as a black and odorous water body， and the model accuracy meets the requirements of black and odorous water body recognition.

Snow depth is an important physical variable in global energy balance and climate change， and accurate snow depth parameters are crucial for global and regional climate and hydrological studies. Active microwave remote sensing has high spatial resolution and is suitable for basin-scale snow depth inversion. As one of the key technologies of active microwave remote sensing， Synthetic Aperture Radar （SAR） can image regardless of weather conditions. However， early SAR systems， while offering high spatial resolution， had low temporal resolution， which made it impossible to perform time-series inversion of snow depth.With the development and launch of new generation SAR satellites，there has been a significant improvement in temporal resolution，providing support for time-series analysis of snow depth. In this study， we selected high-resolution Sentinel-1 data， extracted the phase discretization index threshold， combined with the optical image and high coherence coefficient area， and explored a time series snow depth inversion method based on D-InSAR technology， which successfully inverted the distribution of snow depth in the Wusu area of the northern slope of Tianshan Mountain in the snow accumulation period of 11 days.Sources of snow depth estimation errors are explored based on daily measured snow depth data from three meteorological stations.The study demonstrates that relatively accurate snow depth inversion results can be achieved by employing a phase discretization index threshold extraction method， in conjunction with optical imagery and high-coherence areas for correcting the unwrapped phase.is 0.93， the Root Mean Square Error （RMSE） is 3.98 cm， and the Mean Absolute Percentage Error （MAPE） is 25.49%. Due to differences in interferogram pair coherence and internal properties of the snow， the accuracy of the inversion results was higher when the snow was shallow， with most inverted snow depths being lower than the measured values. Large errors began to appear when the station-observed snow depth exceeded 17 cm， with the maximum error being approximately 7.3 cm.An analysis of the differences reveals that the snow depth inversion accuracy is significantly affected by the differences in image-pair coherence and the actual snow depth. In addition， the inconsistency of the temporal resolution between the optical image and the SAR image may also be one of the factors contributing to the error in snow depth inversion.This method can provide a good estimation of time-series snow depth using SAR data， and reference for “D-InSAR based time-series snow depth inversion”

Large-scale Earthquake disasters often result in significant losses and even casualties. Making prompt assessments of the disaster situation is crucial in the aftermath. Synthetic Aperture Radar （SAR） possesses advantages such as all-weather and all-day capabilities， as well as resilience to lighting and weather conditions. Therefore， the use of SAR imagery for change detection has garnered significant attention in various fields， including post-earthquake rescue and damage estimation， flood extent detection， urbanization studies， and coastline extraction. In this context， this paper proposes a deep learning-based earthquake damage extraction method that integrates spatial and frequency domain texture features. The method demonstrates a robust capability to identify collapsed and intact structures. Using the 2023 earthquake in Kahramanmaras， Turkey， as a case study， the region severely affected by the earthquake， this research incorporates both spatial and frequency domain features into the deep learning network for classification. Experimental results show that the proposed method achieves a classification accuracy of 80.98%， significantly surpassing the original image's accuracy of 47.84%. Moreover， this accuracy is higher than using only spatial domain features （73.30%） or only frequency domain features （73.42%）. The proposed method in this study provides fundamental support for post-earthquake disaster assessment and situational awareness.

The Karlik Mountains glacier located in the eastern part of Xinjiang Tianshan is a typical continental glacier， which is extremely sensitive to the response of climate change. Based on Landsat TM， ETM+ and OLI remote sensing images， DEM data and other information， the glacier boundary information was extracted for four periods of 1990， 2000， 2010 and 2020 using a combination of band ratio method and visual interpretation， and the distribution and variation of glacier area in Karlik Mountain in the eastern Tien Shan Mountains and its response to climate change during the past 30 years were studied.The results show that： （1） the glacier area showed a continuous retreat trend from 1990 to 2020， and the glacier area shrank by 28.34 km² with an average annual retreat rate of 0.73%·a^-1， among which， the retreat rate of the glacier end was the fastest after 2010. （2） As the altitude rises， the distribution of glaciers in the study area shows a trend of increasing and then decreasing， with the most glaciers distributed at altitudes of 3 800~4 600 m； the number and area of small-scale glaciers （≤0.5 km²） are increasing， while the area and number of larger-scale glaciers （≥1 km²） are decreasing； glaciers on different slopes also show different degrees of retreat， with the fastest rate of retreat on the east slope. The distribution of glaciers is characterized by more in the west and less in the east， and more in the north and less in the south； glaciers of different slopes also have obvious retreat trends， among which the retreat is the fastest in the range of 30°~35°. （3） A comprehensive analysis of the climate data in the study area shows that the change of glacier area in the study area from 1990 to 2020 is mainly related to the increase of temperature and decrease of precipitation in the period， and the increase of temperature is the main reason for the acceleration of glacier area retreat.

The cultivation and breeding of rice in Hainan， one of the primary tropical regions in China， play a crucial role in meeting the country's demand for this essential food crop. Currently， there are several challenges in monitoring rice cultivation in the tropical region of Hainan， including limited automation， excessive workload， and low accuracy. In this study， we selected Haikou City in Hainan Province as our experimental area. By utilizing high-resolution multi-spectral satellite remote sensing images such as Jilin-1， Beijing-2， WorldView， and Gaojing-1 along with field verification data， we established a comprehensive database consisting of multi-source and multi-scale samples to accurately identify rice planting areas within the tropical region of Hainan. We employed the DeepLab-V3+ convolutional neural network model for training purposes and proposed an intelligent remote sensing interpretation method specifically tailored for identifying rice planting areas within the tropical region. Experimental results demonstrated that our approach achieved an impressive accuracy rate of 81.9% with a recall rate of 86.7% when extracting rice intelligently based on the DeepLab-V3+ convolutional neural network model. These findings highlight that by training a convolutional neural network model using our interpretive sample database， it becomes possible to accurately extract regions where tropical rice is cultivated from high-resolution multi-spectral remote sensing imagery—a methodology that can serve as a valuable reference for future studies on extracting information related to tropical rice cultivation.

Vegetation cover （FVC）， as an indispensable climate parameter， and the spatial and temporal evolution characteristics of long time series FVC can provide data reference for assessing the surface vegetation condition. MODIS-NDVI data were used to estimate FVC using the image element dichotomous model， and the spatial and temporal evolution characteristics of vegetation cover in Shenyang from 2000 to 2020 were explored by using trend analysis and deviation analysis， while multi-scenario simulation prediction of vegetation cover in Shenyang in 2030 was carried out based on land use data in 2010， 2015 and 2020 combined with PLUS model. The results show that （1） in time， the annual average FVC in Shenyang City increases at a rate of 3.14%/10 a，the high and medium-high vegetation cover shows an increasing trend， and the proportion of vegetation improvement area is higher than that of deterioration. （2） Spatially， the high value areas of FVC in Shenyang are mainly distributed in Shenyei New District， Hunnan District and Sujiatun District， while the low value areas are distributed in the five districts and the central part of districts and counties in the city. （3） The simulation results found that： in the historical trend scenario， the area of arable land， forest land， grassland and water area decreased； in the arable land protection scenario， the area of arable land increased and forest land decreased； in the low-carbon development scenario， forest land increased significantly. The results of the study provide a theoretical basis for the future formulation of environmental management policies in Shenyang.

Exploring snow's response to surface temperature is crucial for understanding snow cover dynamics. In the central Tianshan Mountains， using daily cloud-free snow depth and surface temperature data from 2010~2019， we analyzed coupling， coordination， and lag times via coupled models. Key findings： ①Annual coupling and coordination vary spatially with altitude （rising-declining-rising）， and seasonally （decreasing from winter to summer）， displaying distinct vertical patterns. ② Over the decade， coupling and coordination fluctuated， with increases in the east， decreases in the north， and significant declines below 1 600 m，contrasting with slight gains above the snowline. ③Lag times of snow depth response to temperature increased from spring to winter， peaked during ablation seasons at higher altitudes， and exhibited yearly trends of rising in spring and slight declines in autumn， winter， and summer.

Alpha approximation method is based on the time-invariant of vegetation and surface roughness that receive a great accuracy in soil moisture retrieval. However， the errors may transfer and accumulate as the extension of time scale， and the selection of different prior-information， the retrieval accuracy of this method need to be under reconsideration. This study is based on Sentinel-1 images and carried out in the Tianjun soil moisture observation network that is aimed at the conditional constraints of soil moisture retrieval through Alpha approximation method. The results indicate that： （1） As the ground measurement is used as prior information， the Root Mean Square Error （RMSE） are 0.061 m³/m³， 0.077m³/m³， and 0.090m³/m³ for the monthly， quarterly， and yearly retrieval of soil moisture through Dobson dielectric model. The error is increase with the extension of time scale. （2） As the SMAP product is used as prior information， the RMSE are 0.088 m³/m³， 0.088m³/m³， and 0.101m³/m³ for retrieved soil moisture from the same retrieval strategy， the error increase compared to the results from the ground measurement using. Therefore， the retrieval accuracy is influenced by the quality of prior information. （3） The accuracy of soil moisture retrievals based on Dobson dielectric model and Topp dielectric model is similar in this paper， the difference of RMSE between the retrievals of soil moisture is lower than 0.005 m³/m³. However， the combination of Alpha approximation method and Topp dielectric model can easily extend to the soil moisture retrieval of surface scale.

Timely and accurate acquisition of rice growth information plays an important role in guiding agricultural development and formulating agricultural policy. In this paper， based on the MODIS and Landsat8 remote sensing image data of Sanjiang Plain from 2014 to 2022， the Google Earth Engine （GEE） cloud platform was used to extract the rice planting area and analyze its change trend. Three characteristic indexes （NDVI， EVI and LSWI） were objectively weighted by coefficient of variation method， and Rice Growth Index （RGI） was constructed.The interannual difference model was used to evaluate the rice growth in 2022 compared with the perennial （2014~2021）， and to analyze the spatio-temporal variation and yield trend from tillering stage to filling stage in the Sanjiang Plain. The results showed as follows： （1） From 2014 to 2021， the rice planting area in Sanjiang Plain showed an overall increasing trend， with an increasing area of more than 5 000 km². （2） From June to August 2022， the rice growth in all cities and counties in Sanjiang Plain showed a steady upward trend of "equal tillering stage， best jointing booting stage， and better heading and filling stage"， and the growth in the eastern and northern parts was better than that in the western and southern parts， and the northeastern part showed the best growth. （3） The change trend of rice yield in the 2022 growing season is inferred to be "poor harvest in some areas， high harvest in most areas， and the overall trend of steady increase". The results can provide a scientific basis for agriculture department to guide agricultural activities and domestic rice growth monitoring and yield estimation by remote sensing.

Soil Moisture （SM） is an important parameter for grape development， and accurate monitoring of SM in grape-growing areas is beneficial for scientifically guiding field irrigation and other measures， thereby promoting grape quality and yield. We took Jingyang County， Shaanxi Province as a typical study area， calculated six remote sensing indicators， including Normalized Difference Water Index （NDWI）， Vegetation Supply Water Index （VSWI）， Temperature Condition Index （TCI）， Crop Water Stress Index （CWSI）， etc.， based on MODIS reflectance， surface temperature， and evapotranspiration products from 2009 to 2016， preferentially selected indicators through sensitivity analysis of remote sensing indicators on SM， and constructed a comprehensive remote sensing model of SM using the classification regression tree （CART） algorithm to achieve 8-day resolution scale estimation and drought monitoring of SM at key growth stages of grape. The results showed that NDWI， VSWI and CWSI could better characterize the temporal and spatial variations of SM in the grape-growing area， and NDWI has the most timely response to SM changes and the highest correlation. The comprehensive estimation model of SM constructed using the three preferred NDWI， VSWI and CWSI indicators achieve an R² of 0.6 or more for all time phases except 121 and 169， and an R² of 0.8 or more for all time phases at mid-ripening stage. From 2009 to 2016， the drought in the grape-growing area of Jingyang County occurs mostly during the flowering stage and mid-ripening stage of grapes. This study can provide reference for drought monitoring in grape-growing areas.

Obtaining the boundaries of glacial lakes quickly and accurately from massive remote sensing data is crucial for their inventory. To achieve this， an automatic extraction method based on remote sensing data is needed. This paper presents an improved instance segmentation model based on the YOLOv5-Seg network， which was applied to the automatic extraction of mountain glacial lake boundaries. The results demonstrate that the use of Coordinate-Attention （CA） enhances the network's attention to the glacial lake area. Additionally， a small target detection layer was added to the original three detection layers to improve the network's ability to detect small-area glacial lakes. By modifying the nearest neighbor upsampling method to the deconvolution upsampling method， the upsampling loss feature is solved. Combined with the transfer learning method， this approach reduces the cost of manual labeling. On average， the improved YOLOv5-Seg network achieves an accuracy that is 2.7% higher than that of the original network， reaching 75.1%， and 10% higher than that of other mainstream algorithms. Using the improved instance segmentation model of the YOLOv5-Seg network and Sentinel-2 satellite images， 10 668 glacial lakes were identified in the Hindu Kush-Karakoram-Himalayan region （HKH） in 2022， with a total area of 768.3 km². The study provides the technical basis for the automated mapping of glacial lakes for large geographical regions through the integrated capabilities of deep convolutional neural networks and multi-source remote sensing data.

The accurate and rapid acquisition of soil moisture plays an important role in monitoring， forecasting and warning of regional drought and flood disasters. The high-frequency observation feature of geostationary meteorological satellites provides an effective method for real-time acquisition of large-scale soil moisture information. The reflectance and brightness temperature data of Himawari-8/9， vegetation indices and brightness temperature indices conducted by Himawari-8/9， geographical data， soil data， vegetation status and spatio-temporal information were taken as input features， and the measured soil moisture was taken as expected output feature. A random forest model of soil moisture over Qinghai Plateau was established， and its accuracy was evaluated through independent site testing and spatio-temporal variation analysis of drought processes. The results showed that， the correlation coefficients of Henan soil moisture remote sensing test field and Huzhu remote sensing drought field in 2022 were 0.899 and 0.740， the root mean square errors were 0.062 and 0.044 m³•m^-3， and the mean absolute errors were 0.048 and 0.035 m³•m^-3. In the Huzhu drought process of July 2021， and the Nangqian drought process of August 2022， the variation trend of estimated soil moisture was consistent with the reality. So， the random forest model of soil moisture can meet the real-time monitoring requirement of soil moisture over Qinghai Plateau.

Glacier is one of the most important freshwater reservoirs. Accurate identification of glaciers and monitoring of glacier changes are of great significance for understanding climate change and water resources management. Based on Landsat 8 images， this paper takes the Karakoram region as the research object， and uses single-band threshold method， snow cover index method， unsupervised classification， supervised classification and U-Net convolutional neural network to extract glacier boundaries. The accuracy of glacier boundary extraction results is evaluated by intersection ratio and confusion matrix. The results show that unsupervised classification and single-band threshold method have serious omissions for surface moraine-covered glaciers and glaciers in shadows， and it is easy to misclassify snow-covered mountains into glaciers. The extraction effect of K-means is the worst， with an intersection ratio of 57.69 % and a Kappa coefficient of 0.57. The supervised classification method has significantly improved the extraction effect of moraine-covered glaciers， but the extraction effect of glaciers in the shadow is not good， and the Kappa coefficient of the extraction results is above 0.70. The snow cover index method can effectively extract the glaciers in the shadow， but it is easy to misclassify the non-glacial areas in the large-scale glaciers into glaciers. The intersection ratio is 74.49 %， and the Kappa coefficient is 0.76. The U-Net convolutional neural network can extract the glacier boundary more completely， and the accuracy is significantly higher than other classification methods. The overlapping area is closest to the ground true value area， and the intersection ratio is 88.57 %， and the Kappa coefficient is 0.90. Although the U-Net convolutional neural network performs well， there are still missing points for very small area glaciers. Subsequent research can improve the accuracy by improving the network structure.

Grassland Fractional Vegetation Cover is an important ecological parameter for evaluating the health status of grassland and monitoring environmental changes. At present， the extraction of Fractional Vegetation Cover at large regional scale is mainly based on satellite remote sensing data， and Unmanned Aerial Vehicle Remote Sensing （UAVRS） data， as a supplementary means of estimating grassland cover from satellite data， can improve the accuracy of model estimation. Based on the UAVRS data and BJ3 satellite data， three vegetation cover inversion methods， namely regression analysis method， pixel dichotomy method and random forest were used to invert and validate the vegetation cover of desert grassland in Otog Banner. The results showed that the best inversion model among the regression analysis models established by the vegetation index was the quadratic polynomial model of Normalized Difference Vegetation Index （NDVI）， with R² =0.752； the R² and RMSE obtained from the random forest model directly using the waveband values of UAVRS data were 0.893 and 0.072， compared with the quadratic polynomial model of NDVI and the pixel dichotomy model， R² is improved by 0.141 and 0.151. Using the UAVRS data and the random forest method， it is possible to quickly and accurately obtain the vegetation cover of the study area on the satellite scale， which can provide support for the inversion of desert grassland vegetation cover in the large region.

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.

Rapid and accurate acquisition of cultivated land change information is of great significance to food security management. This paper aims at the problem that remote sensing semantic segmentation method has many errors and omissions due to insufficient model applicability in large-scale and high-resolution image cultivated land non-agricultural detection. Multiscale Scene Classification-Xception （MSC-Xception）， a multi-scale scene classification method for high-resolution cultivated land images based on Xception， is proposed. The convolutional attention module CBAM is embedded into the output layer of the lightweight scene classification network Xception， which has outstanding performance in cultivated land scene classification， to enhance the model's ability to extract channel and spatial features. At the same time， the problem of low separation degree and rough details of mixed scenes existing in the single-scale scene-level classification in large-scale cultivated land extraction is also overcome. Firstly， a feature fusion method of multi-scale cultivated land scene is introduced to improve the separation degree of mixed scene， and then the boundary constraint of multi-scale segmentation vector is used to achieve the boundary refinement of scene-level classification. Compared with the typical Unet， PSPNet and DeeplabV3+ semantic segmentation methods， this method can better reduce the missed detection of large map spots， and the recall rate and F1 score in the cultivated land extraction experiment of GF-2 images in Qixia District in April 2018 increased by at least 15.1 percentage points and 8.8 percentage points respectively. In the non-agricultural detection of cultivated land in Qixia District from 2018 to 2022， the recall rate of suspicious spots increased by at least 7.16 percentage points.

It is of great theoretical and practical significance to study the temporal and spatial changes of ecological quality and its driving factors in the riverbank of the Yangtze River Basin for the coordinated development of ecological environment protection and economic strategy. Based on three Landsat series remote sensing images in 2000， 2010 and 2020， this paper comprehensively evaluated and analyzed the spatial-temporal changes of ecological quality and land use in the Main Stream of the Yangtze River over the past 20 years through the RSEI and Geo Detector. The results showed that： ① The regional ecological quality in the study area showed a trend of decline first and then increase， and the overall trend of regional ecological environment quality decreased obviously from 2000 to 2010， while the ecological environment quality of some regions was significantly improved from 2010 to 2020.②The spatio-temporal variation of land surface type in the study area was obvious， especially the construction land in each county during 2000~2010.③The factor detection found that the influencing intensity of construction land， forest and grass land and cultivated land on eco-environmental quality was different in each year， but they all had strong explanatory power in the study period.④The relevant interactive detection showed that the interactions among the factors of all the years exhibit dual-factor or nonlinear enhancement， while the q values were all greater than 0.48， which indicates that the change of ecological quality further promoted by the interactions among all the factors. ⑤From 2010 to 2020， the ecological protection and restoration in the Main Stream of the Yangtze River achieved initial results. This study provides monitoring and analysis methods and scientific basis for coordinating regional land resource development and ecological environmental protection under the background of rapid urbanization.

Poyang Lake Wetland is an internationally important lake wetland and is of great significance to the healthy and green development of the entire Yangtze River Basin. However， in recent years， the impact of human activities and global climate change has caused rapid changes in the community structure and spatial distribution of Poyang Lake wetland vegetation， and the ecological functions are at risk of degradation. Therefore， timely and accurate knowledge of the spatial distribution of wetland vegetation communities is of great significance for the protection and restoration of Poyang Lake wetland. In view of the characteristics of seasonal changes in the Poyang Lake Wetland， this study uses the 2019 time series Sentinel-2 images as the data source and proposes a random forest classification method that combines statistical and temporal features based on the GEE platform to classify the Poyang Lake National Reserve Wetland. Typical vegetation communities were extracted. The research results show that： （1） Poyang Lake wetland vegetation is affected by periodic flooding and differences in growth phenology， and the differences in vegetation communities are more obvious in autumn and winter. （2） Sentinel-2 images based on time series can better identify vegetation communities， tidal flats and water bodies， among which time features contribute greatly to the identification of vegetation communities. （3） The overall identification accuracy of wetland vegetation communities based on statistical-temporal features combined with random forest classification algorithm reaches 86.21%， and the Kappa coefficient is 0.83. The method proposed in this article can timely and accurately extract the spatial distribution of wetland vegetation communities， and has good application prospects in wetland research. It can provide scientific acquisition and management of wetland resources in Poyang Lake Wetland National Nature Reserve， evaluation of ecological environment， and Repair provides important data support.

Hongze Lake serves as the primary water supply source in northern Jiangsu and functions as the water storage reservoir for the eastern route of the South-to-North Water Diversion Project. The quality of its water is crucial to ensuring safe water supply and the sustainable utilization of water resources. Secchi Disc Depth （SDD） is an important index to measure water environment quality and plays an important role in water ecosystem. In this study， Hongze Lake was taken as the research area， and a remote sensing estimation model suitable for Hongze Lake SDD was constructed based on Landsat 8 OLI remote sensing data by using field measured SDD data and spectral data. The verified results were the Mean Absolute Percentage Error （MAPE） and Root Mean Square Error （RMSE） as 19.8% and 0.07m respectively. The constructed model was applied to Landsat 8 OLI images of Hongze Lake from 2013 to 2022， and the following conclusions were obtained： （1） The average inter-annual variation range of SDD was 0.18~0.25 m， with the highest value and the lowest value appearing in 2019~2020 and 2013~2014， respectively. The overall variation trend of the three lakes was consistent， with the highest SDD in Chengzi Lake Bay and the lowest in Huaihe Lake Bay. The main factors affecting the SDD of the lake area are wind speed， among which the sediment discharge is the main influencing factor of the Huaihe River Bay. （2） The monthly variation of SDD increased from January to August， and decreased month by month after August， with the highest value of 0.36 m in August and the lowest value of 0.20 m in May. The SDD of the lake area is mainly affected by wind speed. Chengzi Lake Bay and Lihe Lake Bay are obviously affected by meteorological factors， and Huaihe Lake Bay is affected by sediment discharge and shipping， and the factors are complicated.

Mitigating climate change is facilitated by the significant carbon reservoir formed through the accumulation of aboveground biomass in secondary forests. The spatial heterogeneity of Guangdong Province significantly affects the carbon sink rate of secondary forests. However， the driving factors remain unclear， which seriously constrains the accurate estimation and prediction of future carbon sink capacity. Based on the advantages of remote sensing technology in spatial and temporal monitoring， this study quantified the impacts of four major drivers on the rate of accumulation and spatial pattern of aboveground biomass in secondary forests in Guangdong Province， using high spatial and temporal resolution remote sensing data on aboveground biomass， year of planting secondary forests， and climate， and combined with scenario analyses to predict the future potential of secondary forests in Guangdong Province in terms of carbon enhancement. The results showed that， overall， stand age was the most important factor influencing aboveground biomass accumulation in Guangdong Province， however， the contributions of other influencing factors were highly spatially heterogeneous. In the Pearl River Delta （PRD） region， climate was the second most important driver， while in the northern， eastern and western regions of Guangdong， it was topography and geomorphology. The effect of soil element content is generally small in the four regions. Among the four scenarios， the maximum storage scenario could maximise carbon gain of 62.45±2.55 Tg C by 2050.This study can provide scientific reference for sustainable forest management and high-quality development.

Over the past two decades， studies about the spectral invariant theory have been developed rapidly in vegetation remote sensing. The theory has been widely used in vegetation parameter measurement， canopy reflectance modeling， and biophysical parameters retrieval. A review paper titled "Photon recollision probability and the spectral invariant theory： Principles， methods， and applications" was published by the author in “Remote Sensing of Environment” in 2023 （DOI： 10.1016/j.rse.2023.113859）. The current paper provides a comprehensive overview of the background of spectral invariant theory， general principles of the theory， determination and applications of spectral invariants. Recent progresses of the theory are summarized and potential future developments are discussed. A special section is dedicated to the researches made by Chinese scholars. The goal is to provide a synthetic overview of the theory. Some new thoughts about the theory are also given in the paper.（1） The spectral invariant theory evolves from the successive orders of scattering approximation method of the multiple scattering process of photons in vegetation canopy. The theory faciliates the conversion of spectral parameters between different spectral bands and different scales （mainly between leaves and canopy） and provides new means for calculating the directional reflectance， albedo， and fluorescence escape probability.（2） Spectral invariants can be obtained through empirical methods， spectral methods， and structural methods. The spectral method is divided into single scale spectral method and spectral scaling method， and the structural method is divided into the Silhouette to Total Area Ratio （STAR） method， clumping index method， Stenberg method， and approximation method. Different methods can be cross-validated.（3） The spectral invariant theory has been applied in a number of canopy reflectance models. Based on these models， researchers have carried out a large number of inversion studies for vegetation structural parameters and physiological parameters. The principles， methods and applications of the theory can be further explored in the future.（4） New approximation methods for the spectral invariants are proposed and new formulae for the visible sunlit leaf area index and the hemispherical directional area scattering factor are summarized. Chinese researchers have made significant contribution to the development of the theory， especially in calculating the escape probability of the solar induced fluorescence using the theory.

The remote sensing identification of river ice provides important support for ice condition monitoring. River ice index identification methods are core tools in river ice remote sensing. However， there is currently a lack of comprehensive comparative studies on common index identification models across different river types. To address this issue， this study applies five remote sensing index models （RDRI， NDSI， MNDSI， NDWI， and reflectance threshold method） to analyze the threshold stability， accuracy， and applicability of these models across six study areas with different river characteristics in the upper reaches of the Yellow River， covering three river types. The results show that the construction methods of the five remote sensing index models consistently indicate that the spectral characteristics of river ice in visible， near-infrared， and shortwave infrared bands are the most critical foundation for river ice identification. The RDRI index performs best in multiple aspects， with an average kappa coefficient of 0.914 4， and is recommended as the optimal choice for river ice index identification. The NDSI and MNDSI indices can effectively eliminate shallow snow interference by adjusting thresholds. The NDSI， MNDSI， and NDWI indices perform well in the headwater study areas， while the reflectance threshold method， though slightly inferior to the RDRI index in performance， still has certain application value due to its simplicity. Among different river types， the five remote sensing index models exhibit the highest accuracy in straight rivers， followed by meandering rivers， and the lowest in braided rivers.

Global climate change and human activities have led to continued increases in the frequency and intensity of droughts. Recently， drought has become a key factor that affects vegetation growth and diversity， which further impacts agricultural production， ecosystem stability， and socioeconomic development. Therefore， mastering the relationship between vegetation dynamics and drought will help to reveal the physiological mechanism of terrestrial ecosystems and formulate effective management strategies. Here， we used long-term datasets of （2001~2020） Solar-Induced chlorophyll Fluorescence （SIF） and Normalized Difference Vegetation Index （NDVI） to explore vegetation changes and their linkage to meteorological drought （SPEI index） across different vegetation types in the Yangtze River Basin （YRB）. Firstly， the correlation analysis method was applied to obtain the maximum correlation coefficient between SPEI and SIF （NDVI）， and the differences of SIF and NDVI responses to meteorological drought of different vegetation types were compared and analyzed. Then we employed an improved partial wavelet coherence method to quantitatively analyze the influence of large-scale climate models and solar activity on the interaction between vegetation response to meteorological drought. The results show that： （1） from 2001 to 2020， the YRB experienced frequent droughts， with summer dryness and wetness exerting the significant impact on its annual climate； （2） SPEI exhibits a greater association with SIF than NDVI does. （3） NDVI has a longer response time （3~6 months） to drought than SIF （1~4 months）， with cropland and grassland displaying shorter response times and evergreen broadleaf and mixed forests showing longer response times. （4） There is a significant positive correlation between drought and vegetation， with a period of 4~16 months. The teleconnection factors of Pacific Decadal Oscillation （PDO）， El Niño Southern Oscillation （ENSO）， and sunspots are crucial drivers in establishing the interaction between drought and vegetation， with sunspots having the most significant impact. Overall， this study indicates that drought is an essential environmental stressor in disturbing vegetation growth over the YRB. Additionally， SIF has great potential and advantages in monitoring drought and vegetation responses. These findings have reference significance for drought prediction， early warning， and ecosystem protection planning in the YRB.

Graph geometrical deep learning has the advantages of modeling topological relationships of long-range ground objects， and describing the boundary of multiple land classes. Existing studies use Principal Component Analysis （PCA） to achieve effective dimensionality reduction of hyperspectral images， but most of them have poor feature separability， which makes the classification performance unable to be further improved. Therefore， the novel hyperspectral remote sensing image classification algorithm based on Graphics Processing Unit （GPU） accelerated t-distributed Stochastic Neighbor Embedding （t-SNE） manifold learning and localized spectral graph filtering was proposed in this study. On the other hand， considering Graph Attention Network （GAT） solves the known shortcomings of previous Graph Convolution Network （GCN） or its approximations by using the hidden self-attention layer， especially since it is good at efficiently processing graph-structured hyperspectral data. Then， the second novel method combining localized spectral graph convolution filtering and GAT network is presented to classify hyperspectral images. Experiments with real hyperspectral datasets on the Microsoft Planet platform show that the proposed methods not only provide new insights into promising hyperspectral image classification performance， but also demonstrate the importance of combining spatial and spectral information for hyperspectral remote sensing image classification.

Terrestrial ecosystem models are important tools for investigating the complex feedback mechanisms between the global carbon cycle and climate change. However， terrestrial ecosystem models are subject to great uncertainties. Constraining model parameters based on observational data is an effective technical approach to realize accurate modelling of the terrestrial ecosystem models. In order to investigate the ability of different observations and their combinations to constrain the parameters of terrestrial ecosystem models and to improve the understanding of terrestrial ecosystem processes， the assimilation of Carbonyl Sulfide（COS）， Sun Induced chlorophyll Fluorescence（SIF）， and Soil Moisture （SM） data were conducted based on the Nanjing University Carbon Assimilation System （NUCAS）. Results showed that the assimilation of COS， SIF and SM could optimize the parameters related to plant photosynthesis and soil hydrology， and improve the modelling of photosynthesis， transpiration and soil hydrological processes in the model. The joint assimilation of COS， SIF and SM can effectively improve the performance of the model in modelling total primary productivity， latent heat flux， sensible heat flux and soil moisture.

Emergencies such as natural disasters and armed conflicts have become significant research hotspots in the global academic community due to their sudden onset and substantial impact. An important research topic is how to efficiently and accurately assess the effects of these events and monitor the subsequent recovery processes. Nighttime light remote sensing， a dynamic subfield within remote sensing， has gained widespread attention in recent years across socio-economic and environmental studies because of its ability to reflect variations in human activities on the Earth’s surface. This study systematically reviews the application of nighttime light remote sensing data in Chinese and international research from 2012 to 2022， focusing on three key areas： armed conflicts， public health emergencies， and natural disasters. The analysis highlights research themes， methodological approaches， and interpretative frameworks. Findings indicate that nighttime light data are used to capture the impacts of emergencies through changes in human activities， the status of public infrastructure， and unexpected shifts in regional economic development. These data have been extensively applied to assess the spatial extent， severity， and recovery progress of emergencies， providing robust evidence for governments to formulate emergency response and post-disaster recovery strategies. With advantages such as accessibility， relative objectivity， and strong correlations with socio-economic factors， nighttime light remote sensing supports rapid and extensive assessment of emergency impacts. Future research should aim to develop nighttime light datasets with higher spatial and temporal resolution， enhance integration with ground monitoring， socio-economic data， and other Earth observation data， and establish a more comprehensive system for emergency monitoring and assessment. This would provide more scientific and effective support for emergency management.

UAV（Unmanned Aerial Vehicle） with hyperspectrometer is widely used monitoring water quality in remote sensing of environment due to lightweight， flexible， low-cost， and abundant spectral information. But it is different from the satellite monitoring system， sky scattered light has a non-negligible impact on acquisition of remote sensing reflectance by UAV hyperspectral. Our research will conduct a discussion on this issue.Lake Chaohu and seven rivers around the lake are selected as the study area. The study used hyperspectral image of UAV and ASD HandHeld2 to measure the spectral signal of water， reference panel and sky light to research on the influence of sky scattered light on the remote sensing reflectance of water. Using optimized reflectance calculation method predicting SPM concentration.The result shows that the MAPD is 11.00% at a zenith angle of 45°and 17.21% at a zenith angle of 0° by substracting different sky scattered light. The result of SPM retrieval indicates that the hyperspectral reflectance subtracting 0°sky scattered light has a lower RMSE and MAPE between measured value and predicted value，8.89 mg/L and 19.60% respectively.Obviously， different angles of sky scattered light have effects on the hyperspectral reflectance of UAV.Above result shows that model of hyperspectral reflectance subtracting 0° sky scattered light has a better performance to predicting SPM. We should subtract 0° sky light hyperspectral to calculate reflectance when predicting water quality parameters based on UAV.

Impervious surface is an important component of urban surface elements. Knowledge about its spatial distribution can provide a scientific reference for urban development and disaster protection. However， due to the similarity of spectra， it is challenging to accurately obtain the impermeable surface material. Object-based and machine learning methods are applied to extract materials of urban impervious. Based on the aerial visible waveband remote sensing imagery with a spatial resolution of sub-meter， the variables including spectrum， vegetation index， texture and shape properties are constructed. Combining Fisher Discriminant Ratio（FDR） and Recursive Feature Elimination （RFE） algorithms， the final variables for training machine learning model were determined. Machine learning algorithms such as Random Forest （RF）， XGBoost， GBDT， CatBoost and LightGBM were developed to construct impervious material classification models （FDR-RFE-RF， FDR-RFE-XGBoost， FDR-RFE-GBDT， FDR-RFE-CatBoost， FDR-RFE-LightGBM）. The best model was selected and to extract the spatial distribution of impervious materials in the study area by comparing the accuracy and the local spatial pattern of impervious materials of different models. The results showed that， compared with the impervious surface material extraction model constructed using all variables， except for GBDT and LightGBM， the overall accuracy and Kappa coefficient values of the models constructed using the variables optimized by FDR and RFE algorithms on the point scale are improved by 0.933%~1.171% and 1.229%~1.542% respectively. Moreover， the phenomenon of spatial fragmentation of classification results is improved. Combining the verification accuracy at the point scale and the local spatial classification results， it was found that the FDR-RFE-RF model showed the most robust performance （OA=0.926， Kappa Coefficient=0.906）， and the spatial distribution of impervious materials extracted for the whole study area was basically accurately represented the ground truth. From our results， we can conclude that variable selection can improve the robustness of impervious surface material extraction based on machine learning to a certain extent. We can also draw the following conclusion that although the aerial visible waveband remote sensing imagery only contains three bands （R， G， B）， it got a reasonable spatial distribution of impervious materials which verifies the potential of visible waveband imagery in urban impervious material extraction.

Henan is a major maize growing province in China， while the southeastern region of Henan is an important maize producing area within the province. Evaluating land suitability can optimize land use structure and strengthen the effectiveness of land management systems. This article took maize cultivation in the Henan Southeast as the research object， comprehensively considering environmental and regional characteristics， and based on the Analytic Hierarchy Process， constructed an evaluation index system for the suitability of maize cultivation and designed a corresponding Comprehensive Evaluation Index （CEI） to complete the regional suitability evaluation. This study calculated the distribution of the CEI for maize planting suitability in the Henan Southeast， and ultimately extracted data within the dryland range for analysis. It is believed that the overall suitability for maize planting in the Henan Southeastern is relatively high， with 98.21% of the dry land being very suitable or more suitable for planting maize. After analysis， this is mainly due to the complementary terrain and hydrothermal conditions in the study area， which have certain environmental advantages and are conducive to the growth of maize.

Tropical islands are among the most biodiverse regions on Earth， with the rivers that are shorter and more diverse than continental rivers. It is crucial to assess the physical habitat health of tropical island rivers for biodiversity protection， water resources management and economic development. At present， the health of river physical habitat is mostly assessed using monitoring points and quadrats， which do not take into account representativeness， convenience， or safety. It is difficult to accurately and completely reflect the status of river physical habitat， and it is time-consuming and laborious. The physical habitat health evaluation index system of tropical island rivers was constructed applying remote sensing technique， with the major rivers in the Nandu River basin， Changhua River basin and Wanquan River Basin of Hainan Island serving as the research objects. The physical habitat health evaluation of 27 rivers in the three river basins was first and comprehensively assessed using hierarchical analysis and expert scoring methods. The results indicated that： （1）The Nandu River Basin， Changhua River Basin and Wanquan River Basin are mainly made up of forest and farmland ecosystems， with the farmland ecosystems accounting for the largest proportion in the entire basin. （2）The vegetation coverage in the riverbank zone of the entire watershed is generally high， reaching over 70%， and is primarily composed of artificial economic forests. The overall natural shoreline rate is low， with an average rate of 41.93%， and the human activities have a significant impact on the riparian zone. （3）The physical habitat evaluation of rivers in the three major basins predominantly falls into a sub-healthy state， followed by an unhealthy state. Based on the current state of ecological and environmental protection in Hainan， relevant research can provide data support for the accurate implementation of ecological protection and governance in the basin， as well as a reference for setting thresholds for relevant evaluation criteria such as water ecological assessment in tropical islands.

The row structure of crops as a typical periodic feature of the cultivated land surface， and its direction will can significantly affect the results of radar backscatter coefficients and optical reflectivity. The key problem of low extraction efficiency and large demand for computational resources when using the texture features of high-resolution remote sensing images to extract crop row direction， it is difficult to be applied on a large scale. In this paper， the feasibility of using the morphological characteristics of the plot to identify the row direction of crops was verified with the premise that the plot was the minimum research object， based on the YouYi County of Heilongjiang Province as the research area. The algorithm uses a variety of images processing algorithms to calculate the length ratio （aspect ratio） between the long side and the short side of the plot， analyzes the relationships between the row direction of the crop and the direction of the long side and the influence of the aspect ratio of different plots on the recognition rate and accuracy. The results show that： With the increase of plot aspect ratio threshold， the recognition rate of row direction decreased from 82.0% to 34.8% ， and the Root Mean Square Error （RMSE） of row direction recognition decreased from 21.46 ° to 1.78°； Under different aspect ratio thresholds， the average accuracy of Line Segment Detector （LSD） algorithm in identifying crop row direction （determination coefficient R²=0.93， RMSE=9.53°） is higher than that of probabilistic Hough transform （R²=0.81， RMSE=20.80°） ； The approach proposed in this paper can effectively achieve the identification of crop row direction in a large range of farmland plots， which provide a new idea for the research of remote sensing satellite images in identifying crop row direction.

Aiming at the problems of small plant size of degraded indicator species and mixed pixels caused by similar morphological characteristics between grass species， a two-stage classification method based on object detection and semantic segmentation is proposed according to the obtained low-altitude UAV data. Secondly， the segmentation model is lightweight improved. The RepVGG network with structural reparameterization is used to replace the Unet backbone network. The efficient channel attention mechanism ECA is introduced in the coding stage， and the feature extraction ability of the model is improved in the down-sampling link to achieve lightweight feature extraction. The block structure uses the ESE module to avoid the loss of channel information. The improved segmentation model has a good classification effect on the two types of grassland degradation indicator species of Artemisia frigida and Convolvulus ammannii in the typical grassland of Xilinhot. The MIoU can reach 0.91， which is about 0.11 higher than the original Unet model. The experimental results show that the UAV data and the two-stage classification method can classify the grassland degradation indicator species well， and the proposed lightweight improved model has a good effect.