In the field of earth observation， cloud detection is not only an important part in the quantitative application of remote sensing， but also a key step in the application of satellite meteorology. In recent years， remote sensing image cloud detection based on machine learning has gradually become a research hotspot in this field， and a series of research achievements have been obtained. Systematically describes the research progress of remote sensing image cloud detection based on machine learning at home and abroad in recent 10 years， dividing the algorithm models into traditional machine learning model and deep learning model. Moreover， the specific models of two categories are introduced in detail， and the advantages， disadvantages and applications of different models are compared and analyzed. This paper focuses on the Support Vector Machine （SVM）， random forest and other methods in traditional machine learning， and the neural network models in deep learning， including Convolutional Neural Network （CNN）， improved U-Net network and so on. On this basis， the existing problems in the research of remote sensing image cloud detection based on machine learning are analyzed， and the potential development direction in the future is discussed.

Aiming at the problem of low precision of abandoned land extraction caused by complex land cover and broken land， a method of abandoned land information extraction based on multi temporal collaborative change detection was proposed. Taking Luquan District， Shijiazhuang City， Hebei Province as the research area， the Normalized Difference Vegetation Index （NDVI） of various types of cultivated land cover was analyzed by using sentinel 2a and Landsat 7 multispectral images and supported by field samples Based on the classification system of seasonal abandonment， perennial abandonment， winter wheat and perennial garden， a multi temporal collaborative change detection model was constructed to carry out remote sensing monitoring of cultivated land abandonment in the study area. The results show that the classification accuracy of seasonal and perennial abandoned farmland based on Sentinel 2A image is 95.83% and 96.55% respectively； the classification accuracy of seasonal and perennial abandoned farmland based on Landsat 7 image is 91.67% and 93.10% respectively； the seasonal abandoned farmland accounts for 4.7% and perennial abandoned farmland accounts for 7.1% in Luquan District in 2019. This method can quickly and accurately obtain the spatial distribution and area information of cultivated land in the study area， and has good extraction accuracy for abandoned land in different resolution images.

Oil palm is a major economic crop and the area of land converted to oil palm cultivation in the tropics has expanded rapidly. Oil palm has become the world's largest source of vegetable oil and it provides tremendous regional economic benefits. However， the expansion of oil palm cultivation has led to the loss of forests， arable land， and peatland， which has caused severe ecological and environmental problems. Application of 3S （RS， GIS， GNSS） technology is useful for the collection， analysis， and management of spatial information， and is essential for both optimizations of the spatial distribution of land use and sustainable development. This paper analyzes the progress of 3S technology application in oil palm research on the basis of a literature review and scientometric analysis. The factors affecting the precision of oil palm mapping are also discussed. We established that papers describing 3S technology application in oil palm research are based primarily on the study of land cover change， and that scientific institutions and researchers in Malaysia， the United States， China， Indonesia， and the United Kingdom are the major contributors. Currently， the application of 3S technology in oil palm research includes oil palm mapping， oil palm land change monitoring， oil palm tree counting， tree age estimation， aboveground biomass and carbon storage estimation， suitability analysis， yield estimation， pest and disease monitoring， and plantation management. The accuracy of mapping is not correlated significantly with the year of publication of specific literature but is correlated with RS data sources and classification methods. The use of 3S technology in oil palm research is currently dominated by RS， which has been used in diverse fields of oil palm research. GIS technology is used mainly for oil palm land change mapping， suitability analysis， plantation management， and pest and disease monitoring， while GNSS is used largely as an additional tool in pest and disease monitoring and plantation management.

Google Earth Engine （GEE） is a comprehensive application platform that integrates remote sensing image storage and analysis. It can conveniently and quickly call remote sensing images and information extraction. Therefore， GEE has attracted more and more scientific researchers' attention. With the continuous expansion and upgrade of GEE， the system platform has become more and more complex. For ordinary users， it is becoming more and more difficult to quickly understand its architecture and functional algorithms. In response to this problem， this article systematically introduces the technical architecture， data resources， model algorithms and computing resources of GEE， and summarizes the application results of GEE in various fields， hoping to provide GEE users with a quick understanding of the platform Window to help them make better use of the GEE platform to carry out their own application research.

Vegetation phenology information is a key indicator for evaluating climate-vegetation interaction， land coverage， and interannual productivity changes in ecosystems. Traditional phenological monitoring methods are based on visual observation， the monitoring range is limited and requires a lot of manpower and resources. As a new monitoring method in recent years， remote sensing technology has the characteristics of large monitoring range， convenient information acquisition and saving manpower and material resources. Its application has promoted the development of vegetation phenology dynamic monitoring research. Firstly， this paper combs the process of vegetation phenology remote sensing monitoring in recent years， and clarifies the existing remote sensing phenology monitoring system； The remote sensing data sources that can be used to establish vegetation growth curve are summarized， and the application scenarios of different data sources are discussed； The existing curve noise reduction algorithms and application processes are summarized， and the causes of errors in different methods are analyzed； The main vegetation phenology extraction methods are summarized； Finally， the remaining uncertainties in remote sensing monitoring of vegetation phenology， such as data resolution， vegetation phenology stage definition， and monitoring timeliness， were discussed， and the main directions for future research on remote sensing monitoring of vegetation phenology were prospected.

Speckle is an inherent property of SAR image， but its existence seriously interferes with the quality of SAR image and affects the high-quality application based on SAR image， so it is urgent to suppress it. The accuracy of the edge detection model of the traditional AD （Anisotropic Diffusion） filter still has room for improvement， and the noise suppression effect is often limited by the problem that it is difficult to accurately estimate the diffusion threshold. To solve the above problems， a novel AD filter based on Multidirectional Sobel （MSAD） is proposed. MSAD filter is an improved algorithm of SRAD. It builds a new edge detection model based on Multidirectional Sobel templates. Based on this， a new AD diffusion coefficient is established by integrating Gaussian kernel function， which can effectively solve the limitation of traditional AD diffusion coefficient by parameter estimation and improve the accuracy of speckle anisotropy suppression. Three real SAR images are selected for filtering experiments. In experiments， SRAD， DPAD， EnLee， and PPB filters are selected as the comparison algorithms； ENL， SSI， ESI， and M-Index are selected to evaluate the performance of proposed algorithms. Experiments show that MSAD filter can effectively improve the edge detection ability and obtain better speckle suppression effect.

Mineral resources are important production materials for human survival and development， and the monitoring of mine environment is crucial for mineral resources exploitation and protection. Due to the advantages including large-scale， multi temporal and comprehensive， remote sensing technology has become the main means of mine monitoring. Aiming to the requirements of mine development and utilization， geological disasters， ecological environment monitoring and quality evaluation， we systematically summarized data sources， methods and models used in remote sensing monitoring of mine environment. Especially， data sources adopted in remote sensing monitoring of mine have tended to diversify and involve in all aspects of mine monitoring. Along with the rapid development of cloud computing platform and artificial intelligence technology， methods such as big data analysis and deep learning have gradually played an important role in remote sensing monitoring of mine environment， while multi-source data fusion， intelligent extraction of features， three-dimensional deformation monitoring and quantitative inversion are the main problems and challenges.

Threshold segmentation method based on SAR image is one of the commonly used methods for effective extraction of water information. In view of the problem of low accuracy and high noise for water extraction on SAR image by Otsu algorithm， a new method based on Otsu algorithm is proposed using C-band Sentinel-1 SAR as the data source. This method constructs natural exponential function based on dual-polarization data to optimize the histogram distribution of pixels in original Sentinel-1 image firstly， and then combines Otsu algorithm to extract water information from image， at last removes the wrongly extracted hill shade based on DEM. The accuracy is evaluated by using optical images of Landsat 8 as the real water information. The results show that the accuracy of water extraction for the proposed method is superior to traditional Otsu algorithm in the case of different water proportions， accuracy of which increased by about 20—60% while water proportion less than 10%. Moreover， this proposed method has low noise and wide applicability features， which can be used for obtaining water information of large area quickly and efficiently.

Farmland is important for food production. It is thus of great importance to obtain timely and accurate information regarding non-agricultural farmlands for land resource management and policymaking. To investigate the changes of non-agricultural farmlands in Fuzhou over past 30 years， this study extracted the spatial information of farmlands using multi-temporal Landsat remote sensing images in 1989， 2000， 2010 and 2019 based on the Google Earth Engine （GEE） and random forest methods. We then used land transfer matrix， grid element method and geographic detector techniques to analyze the characteristics and driving factors of non-agricultural farmlands changes. The results show that： （1） The GEE platform integrating with random forest is suitable to extract farmlands in cloudy and rainy areas in southern part of China. The overall accuracy of the extracted farmlands is higher than 90%， and the Kappa coefficient is greater than 0.85. （2） The farmlands in Fuzhou has an imbalanced spatial distribution， where the area of farmlands deceases from east to west along time. From 1989 to 2019， the farmland changes mainly occurred at areas with an elevation of 100 m and a slope of less than 10°. The changed farmlands mainly consisted of forestlands and construction lands， in which the western region was mainly forestland， and the central and eastern region was construction land. （3） The natural factors are the prerequisite for the conversion of cultivated land， and the growth rate of urbanization and population data are the main driving factors. Moreover， urbanization rate and the proportion of primary industry growth rate were the factors forming the “fast-slow-stable” pattern of farmland non-agriculturalization.

Based on Google Earth Engine （GEE） cloud computing platform， we collaborate with Sentinel-2 images， WordClim bioclimatic data， SRTM topographic data， forest resources planning and design survey data and other data， and use Random Forest （RF）， Support Vector Machine （SVM） and Maximum Entropy （MaxEnt） machine learning algorithms were used as component classifiers to carry out the study of dominant tree species classification with multi-source features and multi-classifier decision fusion. Two serially integrated and three Bayesian parallel integrated models were constructed by the three component classifiers for determining the spatial distribution of 10 major dominant tree species in Shangri-La region of Yunnan. The classification results showed that the overall accuracy of the three component classifiers was lower than 67.17%， the overall accuracy of the three parallel integration methods was comparable， about 72%， the accuracy of the two serial integration methods was higher than 78.48%. Among them， the MaxEnt SVM serial integration method obtained the best accuracy （OA： 80.66%， Kappa： 0.78）， which improved the accuracy compared with the component classifiers by at least 13.49%. The study shows that the decision fusion method has higher accuracy than the component classifier in dominant tree species classification and effectively improves the classification accuracy of small sample tree species， which can be used for dominant tree species classification in large mountainous areas.

Sentinel-2 is a high-resolution optical Earth observation mission within the GMES （Global Monitoring for Environment and Security） programme， which is renamed Copernicus in 2012， jointly implemented by the EC （European Commission） and ESA （European Space Agency） for global land observation with high revisit capability to provide enhanced continuity of data so far provided by SPOT and Landsat. Copernicus is the most ambitious Earth Observation programme to date. It provides accurate， timely and easily accessible information to improve the management of the environment， understand and mitigate the effects of climate change and ensure civil security. At present， Sentinel-2 is one of the most important data source for remote sensing monitoring and application research， and has been widely used in monitoring natural disasters such as floods，forest fires， landslides， volcanic eruptions， and emergency response and humanitarian crises around the globe，and there are also great potentials in detecting glacier and ice and supporting relief efforts for cryospheric disaster.In this study， the glacier and ice avalanches occurred in Arutso Lake basin in northwestern Tibet and Sedongpu basin in southeastern Tibet in 2016 and 2018 were investigated using Sentinel-2 images and field surveys， and the evolution process of two events were reproduced， which has important reference significance for monitoring cryospheric hazard， emergency relief and management in other mountain regions on the world.Study shows that Arutso No. 53 glacier avalanche completely melted away in July 2018 after lasting for two years from occurrence to final disappearance， while the area of Arutso No. 50 glacier avalanche is 0.58 km2 left on June 22，2021 because of more thickness compared to Arutso glacier No. 53. Four large-scale ice-rock ava lanche and debris flow events in the Sedongpu basin in 2017 and 2018 not only had significant impacts on the river flow， landscape and landform in the basin， but also caused great disasters in the basin and downstream.Two glacier and ice avalanche events were caused by climate warming and local heavy precipitation， acting on specific topographic and geomorphic structure of glacier properties in high mountains. Specifically， Arutso glacier avalanche was caused by climate- and weather-driven external forcing， acting on specific polythermal and soft-bed glacier properties and is an unprecedented large catastrophic instability of low angle mountain glaciers. Glacier and snow melting caused by climate warming and heavy rainfall are main triggering factors for ice and rock avalanche in the Sedongpu basin， which is a typical hazard cascades originating from cryosphere， followed by rock fall， debris flow， dammed lake， and lake outburst flood disaster. It often occurs in the Sedongpu basin and will continue to occur for a long time in the future， and the high mountain ridge covered with ice and snow in the right side of back of the basin is still a high-risk area for ice and rock avalanches in the future.

Accurate extraction of ground elevation and vegetation canopy height is of great significance for the study of topography， ecology and so on. The new generation of Global Ecosystem Dynamics Investigation （GEDI） launched in December 2018 provides an unprecedented opportunity for accurate extraction of ground elevation and vegetation canopy height over large areas. The purpose of this paper is to verify the accuracies of ground elevation and canopy height extracted by GEDI using airborne LiDAR data， and to explore the influence of geographic positioning error， terrain slope， aspect， vegetation cover， azimuth， acquisition time， beam type and vegetation type on the estimation accuracy. The results show that the estimation accuracies of ground elevation and canopy height can be significantly improved by correcting the geolocation error of GEDI data. The main factor that affects the extraction accuracy of canopy height is vegetation cover， followed by slope； while the extraction accuracy of ground elevation is significantly affected by the aspect and slope. Additionally， the results also indicated that the estimation accuracy is high when the vegetation cover is more than 25%， and the accuracies of ground elevation and canopy height are the highest in gentle slope area with slope 0~5°. Overall， this study will provide a basis for the screening and application of GEDI data.

Xisha Islands locate in the tropical zone which frequently suffers from cloud cover. Optical systems are vulnerable to bad weather which results in data gaps or low data quality， resulting in difficulties in tropical surface monitoring. To solve this problem， a study on analyzing Xisha vegetation was conducted based on a low-altitude platform. Multi-spectral images were obtained via the DJI Phantom 4 UAV and four vegetation indices from five spectral bands were derived， including the Normalized Difference Vegetation Index （NDVI）， Grassland Chlorophyll Index （GCI）， Green Normalized Difference Vegetation Index （GNDVI）， and Normalized Green-red Difference Index （NGRDI） to analyze the vegetation growth in North island during May 2020. Combined with key meteorological parameters and Worldview2 optical images， the vegetation growth changes between 2020 and 2018 as well as their potential attribution were analyzed. Results showed that the average NDVI， GCI， GNDVI and NGRDI in North Island were 0.30， 0.84， 0.26 and 0.05 in May 2020， reflecting a low vegetation coverage and health status， which was consistent with the ground monitoring results. In 2020， the index difference between artificially managed and natural vegetated region increased from -23%—15% in 2018 to 15%—40%， indicating that the growth of natural vegetation is significantly worse than that of artificially managed vegetation in 2020 which demonstrates strong environmental stress. Meteorological data in this region showed that from April to May 2020， the average daily temperature and wind speed increased while the cumulative precipitation decreased compared with the same period of previous years， leading to increased soil water deficit. This may be the main reason for the deterioration of vegetation growth. These results demonstrated that DJI Phantom 4 images could effectively and quantitatively reflect the vegetation growth which will greatly support the ecological environmental monitoring over tropical islands.

The central urban area of Tianjin is taken as the research object. Based on the abundant OSM road network data and POI big data， functional area identification is carried out at the fine scale. The road space generated by OSM road network data is used to divide the central urban area of Tianjin into 1960 research units. The density distribution and functional area distribution characteristics are analyzed by combining the POI data with weight assignment. The research results show that： （1） In the distribution of urban function density， except for the concentrated distribution of industrial functions in the periphery of the central city， the distribution of other urban functions shows the characteristics of gradual dispersion from the center to the periphery； （2） In a single functional area， commercial areas and public management and public service areas account for a relatively large proportion， while the other four single functional areas account for a small proportion； （3） Among the mixed functional areas， the mixed functional area mainly composed of business-public management and public services has the largest proportion； （4） Comparing the recognition results of functional areas with the Amap， it is found that the accuracy of the recognition results of urban functional areas is relatively high.

The quantitative relationship between FPAR（Fraction of Absorbed Photosynthetically Active Radiation）and vegetation indices has certain reference value for improving FPAR inversion accuracy and guiding production practice. Based on the three-dimensional radiative transfer model LESS， a module named LESS1D （formally released with LESS though www.lessrt.org） with advantages of simplicity of 1D model and accuracy of 3D model is proposed. Based on this model， the influences of vegetation canopy， coverage and other factors on the relationship between FPAR_green and 6 vegetation indices were explored in random homogeneous scenes and 3D heterogeneous scenes. The results showed that in homogeneous scenarios， NDVI， SAVI and EVI fit FPAR_green best in homogeneous scenarios， while NDVI and RVI fit FPARgreen best in heterogeneous scenarios. In heterogeneous scenes， the fitting accuracy of FPAR_green and vegetation index under different crown shapes is cylindrical > ellipsoidal > conical； When the vegetation coverage is low， the fitting accuracy of vegetation indices to FPARgreen is poor； As the solar zenith angle increases， the relationship between RVI and FPAR_green changes from linear to exponential. Canopy volume and canopy geometry are the key factors affecting the size of FPAR_green with different crown shapes， while leaf aggregation， vegetation coverage and vegetation index type are the relevant factors affecting the saturation effect of vegetation index.

Diseases and pests have become one of the biggest constraints on rice yield. Traditional plant protection technology mainly relies on the vision and experience of plant protection personnel， which is subjective， time-consuming and laborious， and difficult to meet the needs of large-scale real-time monitoring. The development of remote sensing technology provides a large-area， all-weather， multi-directional data acquisition method， which can provide crop and environmental information for the identification and classification of diseases and pests， it is an important means to monitor and forecast rice diseases and pests in a large area. On the basis of expounding the mechanism of remote sensing monitoring and prediction of rice diseases and pests， this paper summarizes the research progress of rice diseases and pests monitoring and prediction from many aspects， such as multi-scale remote sensing monitoring method， forecasting method， construction of rice disease and pest monitoring and prediction models， monitoring and forecasting system， etc. ， the existing problems and future development trends of rice disease and pest monitoring and prediction are prospected. With the development of information agriculture and the fusion of multi-source data， the accurate and intelligent remote sensing monitoring and forecasting of rice diseases and pests will become more and more mature.

Sand and its surrounding vegetation types play an important role in fixing dunes， preventing soil erosion and environmental management for sandy land. Identification of Sand and its surrounding vegetation types can objectively reflect the vegetation growth environment of sandy land and its surrounding areas， so as to provide a valuable reference for ecological restoration and the control policies formulating of sandy land. With huge amount of long-term earth observation data and powerful cloud computing capabilities， Google Earth Engine （GEE） cloud platform provides a convenient way for the identification of vegetation types in a large areas. In this study， based on the Sentinel-2 time series data of 2019 stored in the GEE cloud platform， the applied potentialities of GEE cloud platform in vegetation types identification was explored by combining the RF algorithm and vegetation phenology information in Hulunbuir sandy land and its surroundings. Results showed that： ① The spectral information of Sentinel-2 image and the texture information obtained from the near-infrared band have limited ability to identify the main vegetation types in the study area， but the phenological characteristics effectively make up for this shortcoming； ② Accuracy of the vegetation types identification method achieved by the RF algorithm and considering the phenological characteristics extracted from the long time series remote sensing data is 84.37% （with the Kappa coefficient of 0.8）， which is 10.01% higher than that identification result acquired based on single-phase data； ③Phenological characteristics of the main vegetation types in the Hulunbuir sandy land and its surroundings show significant differences， which is helpful for the identification of the vegetation types， especially to improve the recognition accuracy of shrubs and grassland.The research shows that the use of Sentinel-2 data and GEE cloud platform to identify vegetation types in large areas such as sandy land has great potential and broad application prospects.

To explore the application potential of airborne point cloud and UAV visible light image in tree species identification and classification， a single-tree scale tree species classification and recognition method based on UAV hybrid fusion of multi-modal features and decision was proposed. Firstly， Kendall Rank correlation coefficient method and Permutation Importance （PI） were used for feature selection， and Efficient Low-Rank Multi-Mode Fusion Algorithm （LMF） was used to fuse the selected point cloud and visible image features. Ensemble learning was introduced to input point cloud， image， and fusion features into eXtreme Gradient Boosting （XGBoost）， Light Gradient Boosting Machine （LightGBM）， and Random Forest （RF） base classifiers integrated by Stacking. Finally， the meta classifier， Naive Bayes， is used for decision fusion. The experimental data show that the independent test accuracy of the proposed algorithm is 99.4%， which improves 22.58% compared with the Random Forest classifier by traditional feature concatenate fusion. In addition， the Kappa coefficient also increased by 28.54%. The comparison experiment with Convolutional Neural Network（CNN） shows that the proposed algorithm has obvious advantages in small sample training and better generalization ability.

The traditional application mode of remote sensing satellites is complicated and complex， and unable to meet the demand for real-time remote sensing services that users are increasingly concerned about. To equip satellites with intelligent brains can reduce the data transmission bandwidth on the one hand， and improve the time-effectiveness of data acquisition on the other hand. Therefore， on-board intelligent processing has become an essential choice for the development of remote sensing satellites. However， it is difficult to debug on-board processing in orbit， and the existing ground test systems for remote sensing satellite on-board processing platforms are all formed temporarily for different satellite payloads during the satellite laboratory testing. They lack versatility and have not formed an integrated device， resulting in low efficiency of existing ground test platforms for on-board intelligent processing. Especially in the face of the new demand for intelligent on-board processing at present， there is a lack of an on-board processing ground simulation system with high performance， low power consumption and full process. Aiming at the new characteristics of automation and intelligent of remote sensing data processing development， this article proposes a set of ground simulation system for remote sensing image on-board processing based on the combination of FPGA and GPU. This system can realize the 0 to 1 level data pre-processing of multiple payloads in ground simulation and realize the accelerated recognition of intelligent remote sensing images on the basis of pre-processing. The key difficulties are as followed： the balance between the high computational complexity of remote sensing image intelligent processing algorithms and the limited computational power of embedded computers； the balance between the solidification of AI-specific algorithm and hardware acceleration in remote sensing image processing field； the balance between the testing requirements of different satellite platforms and the generality of system architecture. This article illustrates the approach of simulation platform design， builds a basic prototype and verifies it. The test results show that the system can better complete the whole ground testing process for typical algorithms of on-board intelligent processing， and all hardware can be directly assembled on-board with high completeness. It has a certain reference value for optimizing and guiding the operation management system of satellite ground simulation system.

Rice is one of the main grain crops in China， and rice yield is related to people's well-being. Timely and accurate acquisition of rice planting area information and its spatial distribution is of great significance for regional agricultural development planning and yield assessment. To solve the problems of rice mixing easily with other crops and optical data being susceptible to cloud and rain weather， taking the Sanjiang Plain in northeast China as an example， a complete rice phenological growth curve was constructed by using time-series water index SDWI and vegetation index NDVI， respectively， based on sentinel-1 microwave data and Sentinel-2 optical data. The spectral differences of four important growth stages of rice were analyzed， including transplanting stage， tillering stage， heading stage and maturity stage， and the planting area of rice in different phenological stages was extracted by threshold segmentation and combination of data of different stages， and compared with the traditional method based on single optical data. The results show that the proposed method can accurately extract the planting area of rice in several key growth stages in Sanjiang Plain， and is superior to the method using optical data alone. At the same time， the overall accuracy of rice area extraction from single growth period images such as transplanting period images can also reach 87.08%. With the completeness of growth period data， the overall accuracy of rice area extraction based on the whole growth period is also as high as 91.88%， and the Kappa coefficient is 0.834， which can meet the requirements of practical application. Therefore， the rice planting area extraction method combined with multi-source data can accurately and efficiently extract the rice planting area in different phenological periods in Sanjiang Plain， and provide a basis for short-term agricultural situation investigation and management and regional agricultural sustainable development.

Backpack Laser Scanning （BLS） is a potential tool in forest resource survey， but shows much uncertainty for the extraction accuracy of single-tree volume and forest stand volume in complex topographic circumstances. Using BLS point cloud data from the Gaofeng Forest Farm in Guangxi Province， this study implemented the estimation of single-tree volume and sample plot volume by random forest approach. First， individual tree segmentation was conducted using the BLS point cloud data， 8 characteristic parameters were extracted including Diameter at Breast Height （DBH）， Tree Height （Htree）， Crown Diameter （CD）， Crown Area （CA）， Crown Volume （CV）， Canopy Cover （CC）， Gap Fraction （GF）， and Leaf Area Index （LAI）， and 56 stratification height indicators were calculated （height percentage， cumulative height percentage， coefficient of variation， canopy undulation rate， etc.）. Then， an individual treee volume estimation model was developed using the random forest technique， and the prediction accuracy of various parameter combinations was investigated. The results showed that： （1） modeling with only 8 characteristic parameters of an individual tree structure indicated an estimated accuracy of R2=0.83、RMSE=0.097 m³； （2） modeling estimation accuracy was improved with the addition of the layered height index： R2=0.87、RMSE=0.087 m³； （3） the Boruta algorithm for variable screening reduced the input parameters from 64 to 52， with little difference in estimation accuracy： R²=0.87， RMSE=0.087 m³； （4） the estimation accuracy of sample plot volume was R²=0.97， RMSE=0.703 m³·ha^-1. The results suggested the application potential to use the BLS point cloud for individual tree volume estimation and the sample volume by random forest algorithm.

In order to realize fast and accurate extraction of winter wheat planting spatial information by using high-resolution remote sensing image and deep learning semantic segmentation model， worldView-2 remote sensing image was used as the data source to produce the sample data sets with the scales of 128×128， 256×256 and 512×512， which trained the parameters of U-net and DeepLabv3+ semantic segmentation model to establish remote sensing classification model of winter wheat. The classification effects of deep learning was tested by comparing with maximum likelihood and random forest methods. The results showed that： （1） the overall accuracy and Kappa coefficient of the models obtained by training samples of different scales were more than 94% and 0.82， and the model accuracy was stable， which indicated that the sample sizes have little influence on the semantic segmentation model of winter wheat classification. （2） The overall classification accuracy and Kappa coefficient of the deep learning methods were above 94% and 0.89， while the maximum likelihood and random forest were below 92% and 0.85， respectively. This results suggested that the remote sensing classification model of winter wheat established in this study was superior to the traditional classification methods. The results can provide the references for the deep learning methods of crop planting information extraction with high resolution remote sensing image.

The surface Evapotranspiration （ET） is an important controlling factor to water cycle and energy transmission in the biosphere， atmosphere and hydrosphere. Satellite provides an unprecedented spatial distribution of ET in the past decades. In this paper，the estimation methods of evapotranspiration using remotely sensed data were summarized，and the existing issues that should be further studied were discussed. In the future research，we should strengthen the improvement of the evapotranspiration regarding scale effect， nighttime ET， the general validation method of different ET products， remotely sensed data in China， the ET products with higher spatial-temporal resolution， and the new ET model using the machine learning methods.

The carbon budget of terrestrial ecosystems is an important indicator of global carbon cycle research and an important parameter of climate change. Based on the terrestrial ecosystem flux observation and remote sensing satellite observation data， machine learning methods are applied for carbon budget estimation. In this study， random forest algorithm is established to automatically learn features from training data and differences in time series dependencies， and carbon related parameters （Gross Primary Production， GPP； Net Ecosystem Production， NEP） could be estimated. Finally， standard indicators are selected to objectively evaluate the model using the validation data set. The result analysis shows that compared with MODIS GPP products， this method has greatly improved the estimation accuracy. Among them， the prediction result of deciduous broad-leaved forest is the best， the decision coefficient R² is 0.82， and the root mean square error is 1.93 gCm^-2 d^-1.It is also significantly better than traditional light energy utilization model products in other vegetation types. The NEP machine learning model established based on the same method has also obtained good estimation results. The correlation between the output results of the deciduous broad-leaved forest model prediction model and the NEP obtained by the flux tower is 0.70 and RMSE=1.75 g C m^-2 d^-1. The difference in accuracy between GPP and NEP models indicates that when machine learning modeling is performed， the selection of independent variables in the training data set still needs to consider theoretical model. In order to quickly estimate the carbon budget of the terrestrial ecosystem， a remote sensing monitoring platform is established. The platform uses the GEE （Google Earth Engine） big data platform as the data storage and computing backend， and Django， HTML， CSS， JavaScript， etc. as the front-end， in order to quick calculation， real-time visualization and other functions. Based on the platform and algorithm， the global （60° N—60° S） GPP results obtained from 2002 to 2016 show that there are obvious spatial differences in the global average GPP， and the significant increase is mainly concentrated in eastern Asia and forested areas in North America. Research shows that remote sensing monitoring of carbon budget parameters based on machine learning and big data platforms can quickly provide regional and global-scale carbon storage and the results are consistent with true ground observations. The obtained estimation results avoid the complicated parameter setting of the physiological process model， and reduce the uncertainty of regional and global large-scale carbon budget monitor.

Landslides are powerfully explosive and destructive， and are one of the natural disasters with high frequency in the world， causing serious damage to people's lives and properties. Accurate and rapid extraction of landslides and obtaining the distribution range of landslides after a disaster are extremely important for landslide disaster investigation and hazard assessment. The method of landslide monitoring based on high-resolution satellite remote sensing images is investigated. Firstly， the decoding characteristics of landslides on high-resolution satellite remote sensing images are introduced， while the research progress of landslide extraction methods and accuracy evaluation and analysis methods are discussed， and finally the advantages and shortcomings of current methods are summarized， as well as the development direction of future research. The results show that the deep learning method has greater potential， and the combination of deep learning and other automated interpretation methods should be strengthened in landslide monitoring in the future to solve the influence of sample size on the model results， realize the migrability of the model， and improve its automation.

As an important index for monitoring urban ecological environment， the extraction of impervious surface is of great significance. Due to the complexity of urban surface and the need of detailed urban management， it is urgent to extract high-precision urban impervious surface. However， it is difficult to extract high precision impervious surface based on traditional methods. Deep learning method has gradually become a new method of remote sensing image feature extraction because of its characteristics of automatic image feature extraction. Based on this， this paper uses the U-Net deep learning method based on multi-scale feature fusion to improve the semantic segmentation accuracy， and carries out the research on the accurate extraction of impervious surface from high resolution remote sensing images.The residual module is introduced instead of convolution to deepen the network and extract more image features， the pyramid pooling module is added to enhance the network's ability to resolve complex scenarios. It is beneficial to recover spatial information by combining different scale features with jump connection. In this paper， aerial orthophoto images of Guangzhou were taken as the data source. Through convolutional neural network， the remote sensing image is segmtioned into five types of features： background， others， vegetation， road and building. Verify it with the ground truth value of manual visual interpretation， finally， the impervious surface of the study area was extracted. Experiments show that the overall accuracy and Kappa coefficient of the U-NET model are 87.596% and 0.82， respectively. It is superior to traditional supervised taxonomy， object-oriented taxonomy and classical U-Net model in both qualitative and quantitative aspects. The results show that the model can effectively improve the segmentation accuracy of complex scene images by using the multi-dimensional image feature information， and the segmentation effect is good， which is suitable for the extraction of impervious water from high resolution remote sensing images. The research results in this paper can provide data support for urban environmental monitoring.

China has eliminated absolute poverty in 2020， but relative poverty will continue to exist. Doing a good job in the prevention and control of relative poverty is of great significance for consolidating and expanding the results of poverty alleviation and effectively connecting rural revitalization. In view of the shortcomings of poverty classification research and dynamic research that still exist in the existing poverty research， from the perspective of relative poverty， 107 districts and counties in Shaanxi Province are selected as the research objects， and the research period is 2011~2020. Based on night light data， NDVI data and socio-economic statistics data， a multidimensional poverty index estimation model with night light index as the independent variable is constructed to quantitatively identify relatively poor counties， And comprehensively use the analysis methods of Sier index and spatial local autocorrelation to study the spatiotemporal dynamic differences of the relative poverty level in the county quantitatively identified. The results show that： （1） Based on the night light data， the multidimensional poverty index can be estimated effectively， and the estimation accuracy is 84.62%. Using the first 50% of the mean series of Multidimensional Poverty Index as the regional relative poverty division standard is suitable for describing the regional relative poverty level， which is conducive to exploring and establishing a long-term mechanism to solve relative poverty. （2） In terms of time， from 2011 to 2020， the number of relatively poor counties in Shaanxi Province has declined on the whole， and the increase of regional differences in the relative poverty level among cities in the province has directly led to the increase of polarization in poor counties.（3） The relatively poor counties of Shaanxi Province present a pattern of “the degree of poverty is deep and wide in the south， followed by the Weihe River in Guanzhong， and sporadically distributed in the north of the Wuding River in Northern Shaanxi".

Soil moisture is a key parameter in the study of hydrological cycle， ecological environment， climate change， etc.， The acquisition of high-resolution long time series soil moisture information is of great significance for agricultural management and crop growth monitoring， and remote sensing monitoring is also a difficult problem in research. Based on the Sentinel-1 radar data and Sentinel-2 optical data of the time series（2019—2020）， this paper constructs a synergistic retrieval model of surface soil moisture， that is， a method for detecting changes in surface soil moisture under bare soil conditions， And the normalized vegetation index was used to correct the vegetation impact. The proposed method has achieved soil moisture mapping with a spatial resolution of 100 meters in the permafrost region （Wudaoliang） of the Qinghai-Tibet Plateau. The comparison and validation with the in-situ measured soil moisture observed show that the correlation coefficient between the soil moisture estimates and the ground measurements is 0.672≤R≤0.941， and the unbiased root mean square error （ubRMSE） is between 0.031 m³/m³ and 0.073 m³/m³. Soil moisture changes are closely related to regional precipitation events and characteristics， verifying that the change detection method proposed in this study has high applicability in the flat terrain and sparse vegetation areas on the Qinghai-Tibet Plateau.

The existing methods are difficult to accurately estimate the volume of landslides， to solve this problem， the artificial intelligence algorithm is introduced， and transfer learning and differential algorithms coupled landslide volume estimation by low-altitude photogrammetry is proposed. Firstly， high-precision three-dimensional dense point clouds are derived from low-altitude UAV stereo images by using SfM and SGM dense matching algorithms， and ground point clouds are separated from the dense point clouds by combining visible light vegetation index and bilateral filtering algorithm. Then， a deep neural network for data interpolation is constructed to map the nonlinear relationship between two-dimensional coordinates and elevation information， and the elevation value can be predicted based on the transfer learning of parameter sharing and adaptive optimization， and the digital surface model of landslide area can be reconstructed. Finally， the volume of landslide is estimated based on the elevation difference of the digital surface model before and after the landslide in the target area and the differential algorithm. The experimental results show that the average relative error of the proposed method is approximately equal to 2.7%. Compared with the common methods， the proposed method can significantly improve the accuracy of landslide volume estimation， and is suitable for landslide volume estimation under different terrain.

Maize lodging caused by wind disaster may lead to a large reduction in maize production. Using remote sensing technology to accurately monitor maize lodging area and spatial distribution information is very important for disaster assessment.In this paper， Planet and Sentinel-2 images are combined with object-oriented and pixel-based methods to extract maize lodging in the study area， and different image features （spectral features， vegetation index and texture features） and different classification methods （support vector machine SVM， Random forest method RF and maximum likelihood method MLC） influence on the extraction accuracy of corn lodging.The results show that： ① The accuracy of corn lodging extraction using Planet images with high spatial resolution is generally higher than that of Sentinel-2 images；② From the perspective of classification accuracy and area accuracy， the spectral features， vegetation index， and mean feature of Planet image combined with object-oriented RF classification， the overall accuracy and Kappa coefficient are 93.77% and 0.87， respectively， and the average area error is the lowest 4.76%；③The accuracy of extracting maize lodging using Planet and Sentinel-2 images combined with object-oriented classification is higher than that of pixel-based classification. This research not only analyzes the advantages of object-oriented methods， but also evaluates the applicability of using different image data combined with object-oriented methods， which can provide a certain reference for remote sensing to extract crop lodging related research.

Methane is the most representative component of the gaseous hydrocarbon in the marine hydrocarbon seepage. In order to detect the marine methane anomalies accurately，a methane spectra experiment was designed to obtain hyperspectral data of different methane content in seawater background. Based on the measured data， the spectral characteristics of methane are analyzed. The ratio derivative spectrum method is used to weaken the spectral interference of seawater background components for extracting the absorption characteristic band of methane. The results show that methane has spectral absorption in the wavelength range of 1 642—1 672 nm and 2 169—2 378 nm， and the absorption characteristics of methane in the range of 1 642—1 672 nm and 2 169—2 208 nm are significantly enhanced by ratio derivative treatment. Based on the methane index CH₄I， the ratio derivative parameter is added to establish the marine CH₄ content index MI for AVIRIS data. The correlation coefficient R² between MI and methane content is 0.994 2.MI index is applied to the identification of methane anomalies in the hydrocarbon seepage area of the Santa Barbara Channel Coal Oil Point （COP）， California， USA. Compared with the inversion results of CH₄I index and CH₄ index ζ （L2298/L2058）. The abnormal distribution of methane concentration indicated by MI is more consistent with the hydrocarbon leakage area， and the results is better than the inversion results of CH₄I index.

Spatio-temporal fusion of remote sensing images is considered as an effective way to obtain high spatio-temporal resolution data. However， the existing methods require that the low-resolution data at the predicted time is not affected by cloud cover when the basic data pairs is selected， which greatly limits the application potential of the spatio-temporal fusion method. Thus， this article proposes a spatio-temporal fusion method based cloud-covered remote sensing image. Under the deep learning framework， there are two types of remote sensing data featured by high spatial resolution but low temporal resolution （HSLT） and the other type by low spatial resolution but high temporal resolution （LSHT）. The reconstruction subnetwork is constructed to repair the missing information under the cloud coveraged area of LSHT at the prediction dates， and the reconstructed LSHT image and two prior HSLT images are integrated to obtain the final fusion result on the prediction date by the constructed spatiotemporal fusion subnetwork.We take the Landsat （HSLT） and MODIS （LSHT） reflectance data in the coal mining subsidence area of Huainan City， Anhui Province as an example， simulate cloud pollution with different missing rates on the MODIS data at the prediction time， Spatial-temporal fusion experiments are conducted with the proposed method， and then compare water information extraction effects of deep learning fusion data and traditional method fusion data.The results show that the proposed method achieves a good quantitative evaluation effect on the root mean square error and the structural similarity index of the fusion results in each band， and that the fusion results are generally superior to the traditional classical method. The experiment of water extraction in subsidence area clearly shows that the water body extraction result of the proposed method is generally closer to the real observation image. Therefore， the proposed method reduces the data limitation requirements of spatio-temporal fusion， and has higher fusion accuracy and more effective application than the classic traditional method.

Global warming results that glaciers retreat rapidly. Monitoring and mapping glacier boundary are extremely significant for research on global climate change and predicting related disasters. However， snow covering is the main barrier all the time. Selecting Karakoram subregion as study area， the Landsat 8 OLI， and Senitnel-1 images and DEM data in spring （March 24th， 2019） were utilized. The spectral reflectance of green， red， near-infrared and short-wave infrared bands in Landsat 8 OLI images were selected as the optical image features. The backscattering coefficient of VH polarization channel， the coherence coefficient of VV polarization channel， local incident angle， polarization entropy H and scattering Angle α after polarization decomposition were gained from SAR data and used as SAR features. Topographic features included DEM and slope. These characters were employed as input of models. First， based on U-Net model， experiments compared the accuracies using different-size samples. The 256×256-pixel-size samples were imported to U-Net network model based on different backbone networks （MobileNetv2， VGGNet， ResNet and EfficientNet） and DeepLabv3+ model. Finally， the best one among the above networks was employed to import samples with different feature combinations. Results show： ①Using the bigger training sample with the richer spatial context information can obtain the higher segmentation accuracy and the glacier terminal boundary is more accurate. ②Among the different backbone networks， VGG19 backbone network exhibits the highest accuracy， which is higher than that of DeepLabv3+. Its F1-value is 0.899 6， and the mean intersection over union（mIoU） is 0.875 4， and the overall accuracy is 0.948 4. The recognition effect of shadow， snow melt-water， mist covering and frozen lake area is comparatively good. ③With the decrease in the number of training features， the accuracy also drops. Topographic features can improve the precision rate， while SAR features can increase the recall rate by 4% or so. This study proves the feasibility of the deep learning methods on the identification of mountain glaciers covered by a large amount of snow and provides reliable basis on model selection and parameters setting for rapid and large-scale mountain glaciers mapping.

Using the new generation of Geostationary Meteorological Satellite Himawari-8 data， a new adaptive threshold decision tree low-temperature fire spot detection algorithm is proposed. The algorithm is based on the localization recognition results of clear sky pixels and background pixels， using data at channel 2.3 μm and channel 0.86 μm. Taking Shanxi Province as the research area， the identification results were verified using the data on April 24，2020 and February 20，2021. The results of the new algorithm show： （1） the new algorithm can identify small fire spots earlier （sampling point 40 minutes in advance）； （2） the new algorithm has better accuracy on identifying the fire spot with small range and low temperature on grassland and cultivated land； （3） the new algorithm solves the problem of false alarms and misses of fire spot， providing a new idea for identifying fire spot as soon as possible and monitoring fire disaster effectively.

Terrain effect will distort the surface reflectance of remote sensing images， which in turn affects the accuracy of Leaf Area Index（LAI） estimation from reflectance data. To attenuate or eliminate the influence of topography on LAI inversion， a data set of slope reflectance and LAI was constructed as train data based on the three-dimensional radiative transfer model DART （Discrete Anisotropic Radiative Transfer）. With reflectance as the input and LAI as the output， a mountain LAI inversion model was subsequently obtained using random forest algorithm. Then the estimation of LAI in mountain area was realized by combining remote sensing image data in the study area， and the accuracy of LAI estimation was evaluated by the ground measured data. Meanwhile， a flat surface inversion model was constructed based on the DART model and random forest algorithm as a comparison to evaluate the effectiveness of the method proposed. The results indicated that the mountain LAI inversion model considering the influence of topography showed a good performance， with the determination coefficient （R²） of 0.57 and Root Mean Square Error （RMSE） of 0.77 m²/m²， which was better than that of the flat surface model （R² = 0.46 and RMSE = 0.86 m²/m²）.The mountain inversion model based on DART model can capture the influence of slope and aspect on the surface reflectance， and the inversion results can better restore the spatial distribution of LAI in the study area. This study proves that the mountain LAI inversion method based on coupling the DART model and random forest algorithm can partly reduce the terrain effect and effectively improve the estimation accuracy of mountain LAI， which can provide reference for the remote sensing inversion research of vegetation parameters over mountainous areas.

Cloud detection is the basis for related applications using satellite remote sensing images. Aiming at the problem that the cloud detection process is easily disturbed by the complex surface environment， a cloud detection model based on the extreme gradient boosting（XGBoost） is proposed. The method uses Top-Of-Atmosphere （TOA） reflectance， brightness temperature and spectral indices to form a feature space； Then， Bayesian optimization was used to adjust the hyperparameters of XGBoost model. To test the performance of XGBoost in cloud detection， Landsat 8 remote sensing images of different cloud scenes were selected as test data， and the cloud detection results of XGBoost， random forest and decision tree were compared. The results showed that the cloud identification performance of the XGBoost cloud detection model proposed in this paper was better than that of random forest and decision tree， which showed the potential of XGBoost in cloud detection；and the F1 score and Kappa coefficient of XGBoost can reach more than 73% and 71% respectively. The achieved accurate cloud detection and can provide certain support for subsequent researches of cloud detection.

Chengdu-Chongqing urban agglomeration is gradually becoming the growth pole of economic development in western China. Exploring the spatial and temporal pattern of urbanization in Chengdu-Chongqing urban agglomeration has a guiding role for regional coordinated development. Based on the integrated nighttime light remote sensing data from 2000 to 2018， this paper extracted the spatial scope of multi-stage built-up areas of urban agglomerations， using noctilucent scale statistics， standard ellipse， rank-size rule and spatial autocorrelation and other indicators and models to quantitatively analyze the spatial and temporal process of urbanization in this region. The main conclusions are as follows： （1） the multi-year average error of the built-up area extraction with the combination of light and statistical data is 1.27%， which is effective in Chongqing， Chengdu and Mianyang. （2） In the past 19 years， the scale of noctilucent in Chengdu-Chongqing cities increased significantly， with an overall cumulative increase of 5.658 times. After 2010， the scale of light in Chengdu-Chongqing urban agglomeration expanded significantly； （3） The rank-scale of cities in the region shifted from the concentrated development of high-ranking cities to the coordinated and balanced development of the region， and small and medium-sized cities all expanded to varying degrees； （4） The center of gravity of urban agglomeration is located in Anyue County， Ziyang City， Sichuan Province. The center of gravity movement is mainly in the southeast direction， and the spatial pattern evolves along the axis of "Chengdu-Chongqing" from northwest to southeast， indicating that the southeast metropolitan circle dominated by Chongqing has a stronger radiating and driving role and has more influence on the development of urban agglomeration. （5） The spatial agglomeration degree of Chengdu-Chongqing urban agglomeration is gradually strengthened. The overall pattern of cold hot spots is characterized by a large proportion of cold spots and a low proportion of hot spots. The hot spots are mainly located in the main urban areas of Chengdu and Chongqing and their surrounding towns. This study reveals the characteristics and hotspots of balanced development of Chengdu-Chongqing urban agglomeration， which can be used as a reference for future urban function planning and investment decisions.

Rapid and accurate winter wheat acreage extraction using remote sensing technology is of great importance for crop yield estimation and food security. Due to problems such as the difficulty of obtaining medium and high resolution time-series images due to revisit cycles， cloud and rain， and the low accuracy of low resolution remote sensing data in extracting crop planting information. In this study， taking Changge City， Henan Province as an example， Landsat 8 and MODIS images were obtained as the dataset during 2015~2020， and the 2 data were fused based on an optimized convolutional neural network spatio-temporal fusion model to construct a 30 m resolution NDVI time series set， and S-G （Savitzky-Golay） filtering was used to denoise the time series set， and finally The area planted with winter wheat was extracted using the RF method. The results show that the optimised fusion model is robust and the R² of both the predicted and real images is above 0.92. The agreement between wheat area extraction and statistical area in the study area was 97.3% and the results were reliable. Therefore， the optimised model can better fuse the medium and high resolution images， which is an effective technical means to supplement the missing images， and the constructed time series set can more accurately extract the wheat planting area in the county.

FY series satellites can provide important data support for remote sensing monitoring of the atmosphere， land， and ocean on a global scale. As optical satellite images are inevitably affected by cloud coverage， obtaining accurate cloud masks through cloud detection is the key to the processing and application of FY series satellite images. Most of the existing cloud detection methods use simple and efficient threshold methods， however， the optimal threshold in the traditional threshold method is difficult to determine in the absence of a large number of cloud and clear sky labels due to differences in sensor spectral response and radiance differences between different underlying surfaces. Therefore， a Threshold Adapted Cloud Detection （TACD） method is proposed in this paper， which has taken the band characteristics and underlying surfaces differences into consideration comprehensively， then sets up multi-channel threshold tests consisting of reflectance and reflectance combination test， brightness temperature test， brightness temperature difference test and cirrus cloud test under different scenarios， and establish global Optical-LiDAR cloud detection dataset to achieve iteratively optimize thresholds in TACD algorithm， and finally perform cloud detection based on the optimal thresholds. We take FY-3D MERSI-II images as an example to establish a high-precision global cloud detection sample dataset collocated with CALIOP cloud layer data， compare the cloud detection results of the proposed TACD method with the official cloud mask products. The evaluation results show that the accuracy of the cloud masks produced by TACD is significantly improved compared with the official masks， in which the mIoU is increased from 80.35% to 84.09% and the recall can reach 92.67%. In conclusion， TACD has great potential for application.

Hainan province is a golden place to develop tropical characteristic and efficient agriculture. It is of great significance to analyze the change of multiple cropping index with high spatial and temporal resolution. Based on Sentinel-2 data， maximum value composite and Savitzky-Golay filtering and smoothing were used to reconstruct NDVI time series curve. The second difference method was used to calculate the multiple cropping index of cultivated land in Hainan province from 2016 to 2020， and the spatial-temporal evolution characteristics of the multiple cropping index were analyzed. The results showed that the overall accuracy of multiple cropping index extraction in Hainan was 91.94% and the Kappa coefficient was 0.88， verified by the ground survey data in 2020. The multiple cropping index of hainan cultivated land increased from 1.53 in 2016 to 1.66 in 2020， an increase of 0.13. From 2016 to 2020， the single-season planting area increased by 6.10 percent， the two-season planting area decreased by 2.65 percent， the three-season planting area increased by 5.10 percent， and the fallow or abandoned farmland decreased by 5.60 percent. The multiple cropping index of all cities and counties in Hainan province is in the range of 1.28—1.96. The multiple cropping index of Haikou city， Sanya City， Dongfang City， Lingao County increases， while the multiple cropping index of Qionghai City， Wanning City and Qiongzhong County decreases. The results can provide data and decision-making support for agricultural departments in Hainan to adjust fallow and reclamation policies reasonably and implement sustainable development strategy of tropical efficient agriculture.