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.
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.
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.
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.
The basic carrier of grain production is farmland. The rapid and accurate acquisition of non-agricultural information of farmland is of great significance to the management of farmland resources and the implementation of related policies. In order to explore the non-agricultural changes of farmland in Shangnan County in recent 10 years， Google Earth Engine （GEE） was used as the platform， random forest method was used for classification， and temporal and spatial distribution information of farmland in Shangnan County in 2010， 2015 and 2020 was obtained based on multi-temporal Landsat remote sensing images. By means of land transfer matrix， geographic detector technique and grid element method， the important characteristics and driving factors of non-agricultural farmland in Shangnan County were analyzed. The results showed that： （1）The random forest method based on GEE platform could effectively obtain farmland information in Shangnan County， the overall accuracy of land use classification was higher than 88%， and the Kappa coefficient was greater than 0.85. （2） The spatial distribution of farmland in Shangnan County is uneven， mostly distributed in the central and southeastern regions， the farmland area is decreasing continuously， and the non-agricultural development is showing a trend of increasing. The non-agricultural type of farmland is mainly garden land， and the conversion land is mainly concentrated in the central and southeastern areas of Shangnan County. （3） Natural and social factors have a common driving effect on the non-agricultural conversion of farmland， and natural factors are the prerequisite for the non-agricultural conversion of farmland， and the influence of human activities on the non-agricultural conversion of farmland under the limitation of natural factors can be reflected by social factors. The main natural driving factor of farmland non-agricultural transformation in Shangnan County is soil type， and the social factor from road distance has the greatest influence on the transformation of farmland non-agricultural transformation， indicating that the suitable soil type and rapid economic development have a greater impact on the farmland non-agricultural transformation in Shangnan County.
The hyperspectral inversion of soil organic carbon content is of great significance to modern agricultural production and soil quality evaluation. Different soil types have large hyperspectral differences. Exploring appropriate modeling methods for different soil types is conducive to efficient and accurate inversion of soil organic matter content. This study takes the Shanzhou District in the eastern Loess Plateau as the research area and a total of 101 farmland surface soil samples as the research object. Firstly， the spectral data was preprocessed by First Derivative （FD），Cenvelope Removal （CR） and Continuous Wavelet Transform （CWT） respectively. Further， we used Competitive Adaptive Reweighted Sampling（CARS） and Pearson correlation coefficient to filter the feature bands. And finally， we used feature bands as independent variables to compare the three models prediction accuracy of soil surface organic carbon content. The three models were Partial Least Squares Regression（PLSR）、Support Vector Machine （SVM）、Back Propagation Neural Network （ BPNN） respectively. The results show： （1） The characteristic bands processed by CARS are used as independent variables， and the prediction accuracy of the three models （R2=0.67） is greatly improved compared to the model established by extracting characteristic bands as independent variables using the correlation coefficient method， with R2 increased by 0.12； （2） Among the three models， PLSR has the best average simulation accuracy （R2=0.68）， which is significantly higher than SVM （R2=0.53） and BPNN （R2=0.54）； （3） After CWT， the simulation accuracy of models with different decomposition scales is quite different. The CWT26-CARS-PLSR model has the highest accuracy in predicting SOC content （R2=0.91， RMSE=1.29g·kg-1， RPD=3.73）.
In order to improve the overfitting problem of small sample data and improve the generalization ability and prediction accuracy of the wheat stripe rust remote sensing monitoring model， this paper uses the Solar Induced Chlorophyll Fluorescence （SIF） in the canopy obtained by the Chinese Academy of Agricultural Sciences Experimental Station in 2018 as the data source， The Cost Complexity Pruning （CCP） algorithm is used to prune the Random Forest Regression （RFR） method， and the Bayesian Optimization （BO） algorithm is used to select hyperparameter for random forest regression， and a prediction model of wheat stripe rust severity based on the constrained random forest regression （CO-RFR） algorithm is Constructed， And compare the accuracy of the remote sensing monitoring model for wheat stripe rust with the Classification And Regression Tree （CART） algorithm， traditional RFR algorithm， and Multiple Linear Regression （MLR） method. The results indicate that： （1） The CO-RFR model has the highest estimation accuracy and is more suitable for monitoring the severity of wheat stripe rust under small sample data. Among them， in the validation dataset， the average RMSE between the Severity Level （SL） predicted by the CO-RFR model and the measured SL was reduced by 43%， 50%， and 40%， respectively， compared to the RFR， CART， and MLR models， and the average R2 was increased by 56%， 47%， and 40%， respectively. （2） Adding constraints can effectively improve the overfitting phenomenon of the model and enhance its generalization ability. Among them， the average RMSE between the predicted SL value and the measured SL value in the RFR model training set decreased by 62% compared to the validation set， indicating that the accuracy of the model training set was much higher than that of the validation set， and the model showed overfitting. However， the average RMSE between the predicted SL value and the measured SL value in the CO-RFR model training set decreased by 8% compared to the validation set， indicating that the model fitting effect was good and the overfitting phenomenon was significantly improved. eat stripe rust disease under small sample data， and also provides application reference for stress monitoring of other crops.
The solar radiation data plays an important role in land surface energy balance assessment such as sensible heat， latent heat， solar energy assessment etc. However， Radiation data is missing from observed meteorological data by meteorological bureau. Hence， it is imperative to predict solar radiation in a large area with accessible data sources. This research uses the widely accessible data such as （Reanalysis data MERRA 2， remote sensing data MODIS and extra-terrestrial radiation） to drive the commonly used machine learning models for estimating daily solar radiation. The results show that the reanalysis data can replace ground variables and achieve the similar level of precision prediction（difference value 0.14 MJ/m2（MAE）、0.22 MJ/m2（RMSE）、1.13 %（NRMSE））. Particularly， the machine learning models reach the best prediction accuracy（MAE 3.42 MJ/m2，RMSE 4.86 MJ/m2，NRMSE 26.87 %） when driven by re-analysis， remote sensing data and extra-terrestrial radiation together. Meanwhile we also noticed the ensemble of the multiple machine learning model also have a better performance for using any single model. This study highlights that a satisfied radiation data can be generated with widely accessible multi-resource data as well as a couple of the machine learning models.
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 monitoring of river runoff is of great significance to the management and utilization of water resources， but how to obtain river runoff flexibly and accurately is still a difficult problem. Due to the limited resolution of satellite remote sensing， it is difficult to accurately invert the runoff of small and medium-sized rivers. The traditional river flow monitoring technology is complex and expensive， and its application is limited in the areas without data and in the emergency monitoring of sudden disasters. Therefore， this study takes advantage of the fast and flexible characteristics of UAVs and the advantages of LiDAR to obtain terrain information with high accuracy. Based on the 3D model of UAVs LiDAR point cloud data， combined with the Particle Tracking Velocity （PTV） method， This paper presents a method of runoff monitoring for small and medium-sized rivers. In this method， the boundary line between water body and land is extracted by using the strong absorption characteristics of the near infrared band of LiDAR， and the cross section is obtained by matching and merging with the original profile of the river. Based on low-altitude UAV optical remote sensing images， the particle tracking velocity measurement method is used to calculate the river velocity， and then the river runoff is estimated by the velocity area method. After 24 UAV runoff monitoring experiments in the reach of Liancheng Hydrology Station， the following conclusions are reached： The average relative Error between the flow monitored by LiDAR and the measured flow is 8.67%， the minimum relative error is 0.46%， and the Root Mean Squared Error （RMSE） is 0.09 m3/s. MPE （Mean Percentage Error） is 0.02， Pbias （Percent bias） is 1.95%， the Nash-Sutcliffe efficiency coefficient （NSE） was 0.94， which could meet the monitoring accuracy requirements of small and medium-sized rivers in areas without data. By comparison， the monitoring accuracy of runoff using this method is significantly higher than that of Manning formula runoff estimation （RMSE， NSE）. This study demonstrates the feasibility and reliability of the unmanned aerial vehicle Lidar point cloud data runoff monitoring， and provides a new idea for the emergency monitoring of sudden disasters in areas without data and the runoff monitoring of small and medium-sized rivers.
Rapid urban expansion leads to drastic changes in the urban and surrounding ecological environment， which further intensifies the urban heat island （Urban Heat Island， UHI） effect. However， the effect of urban expansion on urban heat island effect in arid areas is still unclear. Gansu province extends from southeast to northwest for more than 1 600 kilometers， and the climate type has gradually changed from humid and semi-humid climate in southeast to extreme arid climate in northwest. Furthermore， Gansu province has experienced rapid urban expansion since 2000， so Gansu province is an ideal experimental area to study the effects of urban expansion on UHI effect under different dry and wet climate backgrounds. This study used the MODIS land surface temperature data set， estimates the surface urban heat island intensity （Surface urban heat island intensity， Is ） and its inter-annual features in 14 cities of Gansu province from 2003 to 2021， to explore the change characteristics of UHI effect and its response to urban expansion in Gansu province under different climatic background. The results show that Isand δISP （Urban-rural contrast in impervious surface percentage， δISP） have obvious spatiotemporal differences among cities in Gansu province. The change trend of Is was mainly affected by vegetation coverage （R2 =0.406， P<0.05）， followed by precipitation （R2 =0.377， P<0.05）， and the effect of urban population （R2 =0.069， P>0.05） was negligible. In addition， due to the difference in land surface temperature （Land surface temperature， LST） among different land cover types， the type of land cover occupied in the process of urban expansion also has an impact on the Is trend. If the type of bare land occupied in the process of urban expansion， it will have a cooling effect. The influence of urban expansion intensity on Is trend has obvious thresholds， and the thresholds of cities in different climate zones are different. The threshold of city in humid zone （Tianshui， δISP=32%） is smaller than that of city in arid zone （Jiuquan， δISP=41%）. The change rate of Is before and after the threshold Is different， and the change rate of Is before and after the threshold Is in arid region Is more significant. The results of this study provide a scientific basis for the impact assessment and management decision-making of rapid urban expansion in arid areas.
The Arctic， as one of Earth's “Three Poles”， is abundant in resources and serves as a key focus in global change research. Changes in sea ice within this region are highly significant for the opening of navigational routes and the preservation of ecological systems. However， existing 3D GIS systems lack the capability to directly visualize irregular grid data， thereby restricting their capacity to support Arctic passage information services. To address this technical challenge， this study proposes a method to automatically transform irregular grid data into regular grid data. Additionally， a Cesium-based three-dimensional visualization system for sea ice data has been developed， enabling the automatic loading and visualization of long-term irregular grid sea ice data within a 3D virtual Earth environment. Performance evaluations demonstrate that the proposed method achieves high levels of accuracy and efficiency， and the developed system effectively and intuitively depicts Arctic sea ice changes， providing critical ice condition information to support ship navigation planning.
Land use/cover change is an essential part of the frontier of global climate change and sustainable development. In this paper， based on Google Earth Engine （GEE） platform， we used Landsat remote sensing images of 1990， 2000， 2010， and 2020， combined with Random Forest （RF） algorithm and field survey to sequentially decode the four phases of land use/cover data in the order of build-up area， water， vegetation， farmland， and unused land and tracked the spatial and temporal patterns of land use/cover changes in the Triangle of Central China. The results show that ① In 2020， the Triangle of Central China will mainly be composed of farmland and vegetation， accounting for more than 86% of the area， with farmland and vegetation mainly distributed around the urban agglomeration， build-up area in a spatial pattern of scattered clusters mosaic； ②Over the past 30 years， the quantitative structure of land use and land-use transfer in the study area have varied， with the land use structure showing an increased continuously in build-up area and farmland， a fluctuating increase trend of vegetation and unused land， and a weak increase trend of water； and ③ The results of the analysis of the evolution of spatial clustering of land use focusing on build-up area and farmland show that the hotter and hot spots of land use degree are increasing. In general， the extreme hot spot area is gradually decreasing， while the hotter sites are on the contrary， and the hot spots are growing in trend， and all of them located in and around Wuhan city circle. The study results lay the foundation for the research on the conservation and sustainable use of land resources in ecological urban agglomeration.
This paper introduces a heterogeneous image registration algorithm called Extended Phase Consistency Spatial Relationship Constraints （EPC-SRC）. The objective is to address challenges related to a high mismatching rate and low registration accuracy encountered in registering heterogeneous images. Initially， we present a chunked Harris feature detection method based on an adaptive weighted moment map. This method aims to identify stable， uniformly distributed， and repeatable key feature points. Subsequently， a multiscale weighted maximum index map （MSW-MIM）， combined with an enhanced GLOH histogram-like （GLOH-like） approach， is employed to construct descriptors. This strategy enhances the distinguishability of descriptors， contributing to improved accuracy. Finally， the Marginalizing Sample Consensus （MAGSAC） method is utilized to establish corresponding points. This facilitates precise feature point matching by incorporating spatial relationship constraints. An iterative approach is employed to solve the homography matrix， resulting in high-precision image registration. A comparative analysis involving ten experiments on aligning heterogeneous remote sensing images demonstrates that our proposed method significantly outperforms other algorithms. Notably， our approach attains superior results in terms of matching accuracy.
Single tree segmentation plays an important role in forest structure analysis， tree parameter extraction and forest biomass inversion. As a low-cost and high-efficiency data source， Light Detection and Ranging （LiDAR） provides a solid data foundation for the study of forest single tree segmentation. At present， the research on single wood segmentation mainly focuses on the forest area with relatively simple structure， and the individual wood segmentation is usually realized by considering the spatial relationship between point clouds and formulating appropriate discrimination criteria. However， for forests with complex structures， there are relatively few existing single tree segmentation algorithms. In this paper， a single tree segmentation algorithm that combines kernel density estimation， digital surface model and K-means clustering methods was proposed. The results show that the method proposed in this study can significantly improve the segmentation accuracy between artificial spruce forest and natural spruce forest when dividing the spruce forest in Northwest China with Gannan Tibetan Autonomous Region of Gansu Province as the study area. Compared with the traditional K-means clustering single tree segmentation algorithm， the overall number of trees in the proposed method is increased by 32% and 15%， and the accuracy is increased by 51% and 27%， respectively， and the recall rates of 83% and 89%， and the accuracy of 92% and 55%， respectively. This method provides a new technical support for the further application of airborne LiDAR in forest ecological applications， especially for the problem of single tree segmentation in complex forest structures.
Aiming at the problems of complex backgrounds， large-scale variations， and uneven distributions in high-resolution remote sensing images， as well as the trade-off between detection accuracy and speed in existing algorithms， a lightweight object detection network called Local-Global Detector （LGDet） is proposed based on YOLO v7. Firstly， the main backbone and neck network were compressed using Partial Convolution （PConv） to reduce the model's parameters and computational complexity. Secondly， a lightweight feature extraction network was constructed by designing a fast Fourier and partial convolution block that combines local and global receptive fields. Lastly， a lightweight triplet attention module was proposed， which enhances useful features. Experimental results on the RSOD and NWPU VHR-10 datasets demonstrate that the proposed algorithm achieves mAP of 93.4% and 90.5% respectively， with FLOPs of 87.9 G. Compared with YOLO v7， the mAP is improved by 2.8% and 2.3% respectively， while reducing FLOPs by 17.3 G. These findings demonstrate that the proposed algorithm achieves superior remote sensing object detection accuracy under lower computational complexity.
Due to the unique natural geographic environment of Ili region， snow disaster is one of the major natural disasters， snow disaster， wind-blown snow disaster and avalanche disaster seriously threaten the operation of regional traffic， and have an important impact on the road planning and other infrastructure construction， the scientific assessment of the risk level of snow disaster on the road is an important basis for ensuring the safe operation of regional traffic and major decision-making such as the location of the road. Based on multi-source remote sensing information fusion data and basic geographic data， this paper develops a set of road snowstorm risk assessment algorithms using entropy weighting and ridge regression methods， so as to construct a potential risk assessment of snow disaster， avalanche disaster and wind-blown snow disaster in the Ili region， as well as a comprehensive risk assessment system of snow disaster， and conducts a comprehensive analysis of the potential risk of snow disaster along the roads in the region. The results show that： （1） the potential riskiness of the three snow disaster types in Ili region has high correlation in spatial distribution， and the potential riskiness of the snow disaster types is in the order of potential riskiness of snowstorms， potential riskiness of wind-blown snowstorms， potential riskiness of avalanches. （2） the comprehensive riskiness of snowstorms basically shows the higher riskiness in the north and south Tienshan region， followed by the central Ili valley region， and the riskiness of small portion of the river valleys and leeward valleys is lower， of which， the riskiness of roads along the routes of snowstorms is also comprehensively analysed. The spatial characteristics of low risk， of which Zhaosu County， Xinyuan County， Nilek County and most of Tekes County are the key areas of snow disaster.（3） from the point of view of the integrated risk of road snow disaster， Zhaosu County， Xinyuan County and part of the road section in Nilek County have the highest risk， followed by the risk of the road section in Horgos City， Huocheng County and Yining County， and the integrated risk of road snow disaster in Gongliu-Xinyuan Junction Road section is lower， and the corresponding road snow disaster risk is higher than that of the other road sections in Ili Valley. The corresponding road sections with higher comprehensive risk of road snowstorms are： G577 Zhaosu-Husongtu Karson section， G577 Zhaosu-Tex section， S330 Kurningde-Nalati section， G218 Nalati-Gongnaix section， and G578 Nilek-Tambura section.
The rural buildings are the most important disaster recipient in the earthquake disaster， it has significant meaning in the fields of earthquake resistance and hazardous reduction to obtain the information like the type and distribution of it. Based on GF-2 high-resolution remote sensing data， the ESP （Estimation of Scale Parameter） algorithm and Seath（Seperability and thresholds） algorithm are used to determine the optimal image segmentation scale and construct the optimal feature learning space. The decision tree classification method and random forest machine learning classification method were chosen to extract and classify rural building structures in Xiangnan Town， Gansu Province，in early May 2021. Unmanned aerial survey and on-site investigation data were used to verify and refine the accuracy of the classification results. The results show that： ①Both methods can better identify brick-concrete buildings with uniform spatial distribution， large area and bright color， but for civil buildings with chaotic distribution and relatively concentrated， gray color and small area （brick-wood buildings） are difficult to effectively identify their boundary contours and accurately classify them. ② The accuracy and Kappa coefficient of the two methods for building classification in the study area are 84.42%， 86.82% and 0.701 5， 0.759 1， respectively， and the random forest-based method has less misclassification and missing phenomenon when extracting building information. Therefore， the random forest method is more suitable for rural building classification.
The next-generation satellite LiDAR systems， including the Ice， Cloud， and land Elevation Satellite-2 （ICESat-2） and the Global Ecosystem Dynamics Investigation （GEDI）， offer unique advantages in estimating forest canopy height. The fusion of these two LiDAR datasets not only increases the sample size for canopy height retrieval but also allows for spatial complementarity between different datasets. First， the Random Forest-Recursive Feature Elimination （RF-RFE） method was used to select photon feature parameters. Subsequently， five fusion models—Stepwise Linear Regression （SLR）， Random Forest （RF）， Light Gradient Boosting Machine （LightGBM）， Random Forest with Stepwise Linear Regression （RF-SLR）， and Particle Swarm Optimization Random Forest （PSO-RF）—were analyzed for their applicability. The optimal model was selected to construct a point-scale canopy height dataset， which was then combined with multi-source remote sensing imagery to map the canopy height in Qilian Mountain National Park. Finally， the retrieval results were compared with existing canopy height products using GEDI footprint data and field survey data. The results showed that： （1） the Particle Swarm Optimization Random Forest （PSO-RF） model provided the best fusion performance （R² = 0.71； RMSE = 3.15 m； MAE = 2.66 m）； （2） the retrieval model based on PSO-RF fusion of point-scale canopy height data achieved the highest accuracy （R² = 0.56； RMSE = 3.02 m； MAE = 2.38 m）； （3） compared to existing canopy height products， the retrieval results demonstrated higher accuracy （based on GEDI footprint data： R² = 0.43； RMSE = 4.50 m； MAE = 3.59 m）， and the errors were smaller when compared to field survey data （R² = 0.36； RMSE = 3.15 m； MAE = 2.56 m）. The findings reflect the spatial distribution pattern of vegetation in Qilian Mountain National Park and provide scientific support for forest resource management， carbon sequestration estimation， and ecological resource conservation.
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.
High resolution optical images are susceptible to imaging conditions and changes in ground features， resulting in geometric and radiative distortions which are difficult to ignore. Traditional registration methods are difficult to ensure accuracy and stability. Therefore， this study fully utilizes the advantages of phase congruency in the geometric structure expression of images， proposes a registration algorithm based on multi-directional phase congruency， uses Log-Gabor filter to calculate the phase congruency and minimum moments of images in different directions， further extracts image corners through multi threshold constraints， and constructs descriptors that are weakly sensitive to radiation and rotation to achieve complex radiation Matching homonymous features in geometric distortion scenarios， and finally using matching features to solve local registration models. Multiple experiments based on GF-1/2/6/7 images have demonstrated that the proposed algorithm can achieve high-precision registration and is suitable for domestic high-resolution optical image registration tasks from different sources， scenes， and resolutions.
As an important vegetation ecosystem， the study of phenological changes in grasslands is important for a comprehensive understanding of global climate change and ecosystem carbon cycling. Inner Mongolia has the most important grassland resources in China. In this study， we employed the double logistic method to extract the key phenological parameters， including the start （SOS）， peak （POS）， end （EOS）， and length （LOS） of the growing season， of Inner Mongolia's grasslands using GIMMS NDVI3g data over a period of 34 years （1982~2015）. Subsequently， we analyzed the spatial and temporal variation of these parameters， as well as their response to meteorological factors and the impact on productivity. Our findings revealed that the SOS in Inner Mongolia grasslands was mainly concentrated from early May to early June， while POS was primarily concentrated from early July to early August. Moreover， the EOS mainly occurred from mid-October to early November， while LOS ranged from 140 to 170 days. We observed that SOS， POS， and EOS were all dominated by the trend of advancement， with the proportions of extended and shortened LOS being relatively similar. In addition， we found that SOS and POS were jointly influenced by temperature and precipitation， while EOS was mainly influenced by precipitation and less by temperature. Our analysis further indicated that an advancement of SOS， postponement of EOS， and extension of LOS would result in a decrease in GPP， while an advancement of POS would lead to an increase in GPP.
It is important to analyze the change of the quantity and spatial pattern of cultivated land for evaluating the current situation of regional food security and sustainable utilization of cultivated land resources. Based on the second national land survey and change data， this paper analyzes the changes in the quantity and spatial distribution of cultivated land in Northeast China from 2009 to 2018 by means of spatial analysis and trend analysis， and reveals the pattern of newly added cultivated land and its physical geographical characteristics. The results showed that： （1） the cultivated land in Northeast China increased slightly from 2009 to 2018， and the growth rate was slower than before； （2） The newly added cultivated land was mainly distributed in the west Liaohe Plain， the west Songliao Plain and the Sanjiang Plain， mainly from the reclamation of forest grassland and unused land； （3） In Northeast China， the water and heat conditions of newly added cultivated land were slightly better than those of traditional cultivated land， but the situation was opposite in Sanjiang Plain and Lesser Hinggan Ling-Changbai Mountains. Although the newly added cultivated land is mainly distributed in the plain area， the phenomenon of cultivated land going up the mountain still appears in some areas. At the same time， it was found that the soil fertility of newly added cultivated land in eastern Inner Mongolia and Songliao Plain was slightly weaker than that of traditional cultivated land. The study suggests that the natural conditions of newly added cultivated land are different from those of traditional cultivated land. Considering the ecological risks brought by cultivated land reclamation， quality assessment and monitoring of newly added cultivated land should be strengthened in the future， especially the risks of economic value-added and ecological degradation of cultivated land reclamation should be weighed， so as to promote the orderly utilization and sustainable development of cultivated land resources.
To improve the accuracy of sea fog recognition， a daytime sea fog recognition method is proposed by using a generative adversarial network with multi-scale feature fusion under the attention mechanism. The method first generated cloud images for the mid-infrared channel using a conditional generation adversarial network to eliminate the solar radiation effects of the original daytime mid-infrared channel cloud images， allowing the visible， far-infrared， and mid-infrared channel cloud images to be comprehensively utilized for sea fog monitoring. On these grounds， the pyramid split attention mechanism was introduced into the UNet network to improve the performance of data feature extraction for the three input channels. At the same time， multi-scale atrous spatial pyramid pooling was used in the transition layer of the codec to enhance the generalization ability of sea fog recognition at different scales by multi-scale feature fusion of multiple paths. Finally， the discriminant network was introduced to supervise the generation network to achieve an accurate definition of the edge of sea fog. The experimental results show that the accuracy of sea fog detection in this method is improved compared with the traditional method， the Probability Of Detection （POD） reaches 94.16 %， the False Alarm Ratio （FAR） is 11.61 %， and the Critical Success Index （CSI） is 83.59 %， which provides a new idea for daytime sea fog recognition.
