Leaf Area Index （LAI） is an important parameter for studying vegetation canopy structure and physiological and biochemical characteristics. Due to the high complexity and heterogeneity of mountain’s surface and forest canopy structure， there is no unified standard for mountain LAI ground measurement methods， resulting in significant differences in ground measurement between different methods. In order to analyze the effect of various factors on the ground measurement， and improve the accuracy and reliability of LAI ground measurement data in mountainous areas， this article uses LAI2200 Plant Canopy Analyzers （LAI2200） and Digital Hemisphere Photography （DHP） to conduct LAI ground observation experiments in typical mountain forest scenes， quantitatively analyzing different optical measurement instruments， vegetation clumping effects， the impact of terrain and other factors on the LAI ground measurement in mountainous areas. The results showed that both LAI2200 and DHP optical instrument could be used to measure LAI in mountain forests. The experiment found that LAI2200 was more sensitive to LAI of coniferous forest than DHP； Regarding the impact of terrain factors on LAI ground measurement， the hinder of incident radiation by surrounding terrain fluctuations is an important factor that causes measurement errors and needs to be eliminated in the measurement； In addition， the vegetation clumping effect has a significant impact on the measurement results， and it is necessary to correct the effective LAI by introducing the Clumping Index （CI） to obtain the true LAI. This article provides effective reference suggestions for improving the accuracy and precision of LAI ground measurement in mountainous forests.
In the process of obtaining agricultural planting structure through remote sensing images， the uncertainty of remote sensing images itself and the classification process will inevitably reduce the accuracy and reliability of mapping results. As the basic unit of agriculture， parcels have the advantages of classification accuracy and uncertainty control. However， current remote sensing classification uncertainty research methods mostly use pixels as the basic unit， which is difficult to directly apply to parcels crop classification. Therefore， using parcels as the basic spatial unit and selects information entropy as the posterior uncertainty evaluation index to carry out crop classification and uncertainty analysis experiments in the Ningxia Irrigation District of the Yellow River. With the help of temporal characteristics and multiple value characteristics inside the parcel， the influence effects of random uncertainty and fuzzy uncertainty in remote sensing data on posterior classification uncertainty are analyzed. The experimental results show that： （1） Compared with pixel-scale classification， parcel-scale classification can effectively weaken posterior uncertainty； （2） Random uncertainty has a significant impact on posterior uncertainty and is significantly correlated with temporal characteristics； （3） The fuzzy uncertainty introduced by mixed pixels at the edge of the parcel and internal mixed heterogeneity amplifies the random uncertainty and further affects the posterior uncertainty. Targeted reduction measures can effectively reduce classification uncertainty， and the reduction effect is basically consistent with the correlation analysis results. The research results can provide ideas and methods for subsequent improvement of crop classification process and uncertainty control.
Accurate estimation of large-scale terrestrial evapotranspiration （ET） is critically important for hydrological， ecological， agricultural， and water resource management. Penman-Monteith （PM） equation coupled with the Jarvis stomatal conductance model plays a pivotal role in remote sensing-based evapotranspiration estimation models. However， current PM-Jarvis frameworks rely on optimization-based methods to derive maximum stomatal conductance （gsm） for vegetation or International Geosphere-Biosphere Programme （IGBP） categories， which lack a biophysical foundation. These approaches overlook interspecific variations in stomatal traits and fail to capture the temporal dynamics of leaf-level gsm， leading to significant estimation errors. This study proposes a novel methodology to calculate gsm using species-specific stomatal anatomical characteristics （stomatal length and density） and a physiological constraint function based on Normalized Difference Vegetation Index （NDVI）， enabling accurate estimation of surface evapotranspiration.. Stomatal anatomical characteristics can determine gsm of various plant types and physiological constraint function can simulate the dynamic changes of gsm. Validation using observational data from sites of AmeriFlux network demonstrates that compared to the approach that does not distinguish vegetation types and based on IGBP fixed gsm， the precise calculation of dynamic gsm significantly improves evapotranspiration estimation accuracy. The coefficients of determination （R2） increase from 0.626 and 0.726 to 0.83， and the Root Mean Square Error （RMSE） decreases from 33.31 W/m2 and 27.72 W/m2 to 21.81 W/m2， respectively. Therefore， a more detailed consideration of vegetation physiological differences in large-scale evapotranspiration estimation proves to be an effective method for enhancing accuracy.
China's independently developed HY-2 series of power satellites provide important data support for global ocean wind field observation and research. By fusing the sea surface wind field data acquired by multi-source remote sensing satellites to form wind field fusion data with high temporal and spatial resolution， it is of great significance to the study of typhoon disasters and the safeguarding of the safety of ship navigation and offshore operations. Cross-validation of wind field fusion data is a prerequisite for the large-scale application of this data. In this paper， the Northwest Pacific Ocean， which is prone to typhoon disasters and rich in fishery resources， is selected as the study area， and ERA5 reanalysis data， CCMP wind field data and CERSAT wind field data are utilized as the reference data. By obtaining the standard deviation and correlation coefficients of wind speed and direction between different datasets， the cross-validation analysis of HY-2 series satellite wind field fusion data between different seasons from December 2019 to November 2021 was carried out. The results show that： ① the HY-2 series satellite wind field fusion data match the ERA5 reanalysis data better than the CCMP wind field data and CERSAT wind field data. ② In different seasons， the wind speed and direction accuracy of the HY-2 series satellite wind field fusion data is higher in summer and fall， and relatively lower in winter and spring. ③ The wind speed and wind direction of the HY-2 series satellite wind field fusion data from December 2019 to November 2021 have high accuracy， with standard deviations of 1.54 m/s and 13.29°， and correlation coefficients of 0.92 and 0.88， respectively， and can be widely used in the Northwest Pacific Ocean.
Surface albedo is a crucial parameter in determining the surface energy balance. Site observations from solar radiometers are essential for acquiring surface albedo data and validating surface albedo remote sensing products. Traditionally， surface albedo site observations have been primarily used to validate low spatial resolution satellite remote sensing products， which leads to significant scale mismatch issues when validating high spatial resolution remote sensing products. Additionally， the footprint of mountainous surface albedo site observations differs significantly from that of flat surfaces due to the influence of terrain. With the rapid development of algorithms and production of medium to high-resolution surface albedo products， there is an urgent need for high-resolution pixel-scale "truth" data in mountainous areas to verify their accuracy. To address these challenges， this study proposes a remote sensing spatial downscaling model for mountainous surface albedo site observations， based on the characteristics of solar radiometer observations， surface reflectance properties， and local terrain features. Firstly， the model simulates the actual observation footprint range in mountainous regions using digital terrain viewshed analysis. Then， it corrects the upward shortwave irradiance observed by the radiometer according to its cosine response characteristics. Furthermore， a decomposition method is developed to account for local terrain features， which uses high-resolution reflectance band integration to characterize pixel reflectance properties. This method decomposes the corrected upward shortwave radiation and the calculated downward shortwave radiation to pixel-scale， ultimately obtaining high-resolution pixel scale surface albedo within the radiometer observation footprint range. This approach enables the acquisition of pixel-scale surface albedo reference "truth" products from site observation footprints. Using a three-dimensional radiative transfer （LESS） model， this study simulates high-resolution pixel surface albedo within the observation footprint range under clear-sky conditions to validate the method.The validation results demonstrate an R2 of 0.83，RMSE of 0.016 9， and a linear fitting slope of 0.999 between simulated surface albedo and decomposed surface albedo， indicating the effectiveness of the proposed model in decomposition and downscaling of mountainous surface albedo of site observation footprints. Furthermore， using real remote sensing observations and site observation data， a high-resolution pixel-scale surface albedo downscaled dataset has been constructed for the Wanglang Mountain Remote Sensing Observation and Research Station of Sichuan Province. This model provides a novel solution and approach for the development and validation of high-resolution mountainous surface albedo retrieval algorithms and products.
Surface soil moisture is a key variable controlling water cycle， carbon cycle， and energy exchange between land and atmosphere. Currently， passive microwave observations in the L-band are considered the optimal wavelength for retrieving surface soil moisture information. However， the low spatial resolution of passive microwave observations is insufficient to meet the needs of applications such as hydrological modeling， weather forecasting， agricultural planning， and water resources management. The direct observation of passive microwave remote sensing is brightness temperature （TB）. Therefore， obtaining high-resolution brightness temperature is the foundation for obtaining high-resolution soil moisture. Addressing this limitation， our study assesses the potential of refining Soil Moisture Active Passive （SMAP） L-band data resolution by integrating it with X-band data from the Advanced Microwave Scanning Radiometer-2 （AMSR-2）. We employed a Time-Series Regression （TSR） approach， incorporating soil and vegetation descriptors such as the Microwave Vegetation Index （MVI）. The enhanced resolution of the downscaling was validated using ELBARA-III TB data， the results show that the precision of TSR-MVI downscaling TB can be consistent with the original SMAP data. The minimum Root Mean Square Error （RMSE） at V-polarization is 9.864 K， and the maximum correlation coefficient （R） is 0.861. The downscaling results based on TSR-MVI method are superior to those of Backus-Gilbert optimal interpolation， with more information and higher image clarity. Our findings suggest that the TSR approach， when combined with AMSR-2 TB data， can effectively downscale SMAP L-band data.This method can downscale the L-band SMAP brightness temperature， and the results are superior to those of the BG product.
To monitor the spatiotemporal variations of soil moisture over the Tibetan Plateau—known as the "Roof of the World" and the "Water Tower of Asia"—in the early 21st century， this study evaluates multiple soil moisture products and adopts the AMSR-E/2 soil moisture dataset retrieved using the Multi-Channel Collaborative Algorithm （MCCA）， which demonstrated the highest accuracy. Based on data from 2002 to 2022， we applied Sen’s slope estimator combined with the Mann–Kendall test to systematically analyze the spatiotemporal patterns of soil moisture changes and its relationship with precipitation across the plateau. The results show that the Tibetan Plateau has experienced a general wetting trend in the early 21st century， with the proportion of significantly increasing soil moisture trends （31.52%） far exceeding that of significantly decreasing trends （1.70%）. Most river basins exhibit a pattern of “high increase in dry areas and low increase in wet areas”.Furthermore， notable differences exist in the trends among various river basins and monsoon zones. Precipitation is identified as a key driving factor influencing soil moisture changes on the Tibetan Plateau.
Soil moisture is a crucial variable in terrestrial ecosystems and serves as a bridge for land-atmosphere interactions. Microwave remote sensing enables the acquisition of spatial distribution information of soil moisture at a large scale， and numerous products have been released. However， there is no single sensor data product that consistently maintains optimal accuracy across all spatial and temporal dimensions. Data fusion represents an effective approach to enhance the accuracy of soil moisture data， and precise analysis of the error structures within the data facilitates the construction of efficient and precise data fusion strategies. This ensures the selection of the most suitable algorithm during the data fusion process and minimizes the negative impact of errors on the fusion results. Taking the Tibetan Plateau as the study area， this paper utilizes three soil moisture products from SMAP， SMOS， and AMSR2 as source data for fusion， and employs soil moisture converted from remote sensing apparent thermal inertia data as the reference data. The Triple Collocation method is applied to decompose the errors in the source data， and the systematic errors are corrected accordingly. Then， based on the random errors of the corrected source data， weights are assigned to obtain a fused dataset. The results demonstrate that the fused dataset exhibits better temporal and spatial continuity and higher accuracy compared to the source data. Specifically， in the two validation regions of Babao and Naqu， the Root Mean Square Error （RMSE） decreases by 0.104 cm³/cm³ and 0.043 cm³/cm³， respectively， and the Mean Absolute Error （MAE） decreases by 0.099 cm³/cm³ and 0.038 cm³/cm³， respectively. This proves that the method presented in this paper performs well in data fusion and can provide valuable insights for related research and applications.
To promote the application of Fengyun satellite data in the real-time monitoring of surface soil moisture over Qinghai-Tibet Plateau， a Hydro-Thermal Condition Index was used to characterize the combined effect of temperature and precipitation in the process of soil moisture change， a soil moisture retrieval model was developed by random forest algorithm. In this model， FY3D/MWRI bright temperature data and its passive microwave indexes， auxiliary data such as soil properties， geographical topography， vegetation status， and Hydro-Thermal Condition Index， were used as independent variables， and soil moisture measurement data was used as dependent variable. Independent inspection and drought process inspection were carried out in Henan validation field and Huzhu drought field of Qinghai Province. The results showed that the contribution of Hydro-Thermal Condition Index was the highest （0.229）. The model had good inversion accuracy， with correlation coefficients of 0.874 and 0.860， root mean square errors of 0.061 and 0.026 m3/m3， and average absolute errors of 0.050 and 0.022 m3/m3 in Henan validation and Huzhu drought field， respectively. And it can effectively capture the dynamic characteristics of soil moisture during the summer drought in the eastern agricultural area in 2023. This study can provide a technical reference for all-weather soil moisture monitoring， drought and flood monitoring based on Fengyun meteorological satellite over Qinghai Plateau.
A ccurately and rapidly extracting tobacco planting distribution information holds significant importance for optimizing crop planting structures and scientifically planning tobacco field layouts. Mianchi County， a key tobacco planting base in Henan Province， features a hilly terrain characterized by fragmented plots and complex mixed-crop patterns. These characteristics make it challenging to meet the demand for precise tobacco extraction in hilly areas using optical remote sensing features alone. Consequently， integrating multi-source remote sensing imagery， selecting optimal remote sensing phases for tobacco classification， identifying significant features for tobacco remote sensing classification， and exploring feature optimization methods are crucial for enhancing the accuracy and reliability of tobacco classification in such regions. This study leverages the Google Earth Engine （GEE） platform and utilizes Sentinel-1/2 imagery to extract spectral， polarization， index， and texture features of ground objects within the study area. Notably， the index features incorporate the red-edge index， derived from red-edge spectral calculations. Employing an object-oriented approach， six classification schemes were designed based on the Random Forest algorithm to investigate the impact of various feature type combinations on tobacco planting distribution information extraction. The research findings reveal the following key insights： Optimal Segmentation Scale： In the object-oriented method using the SNIC algorithm， a segmentation scale of 3 pixels yields the clearest and most complete land class details in the segmented imagery. Feature Reduction via J-M Distance： By applying the J-M distance separability metric， the number of classification features was reduced from 28 to 15， effectively retaining the essential information required for accurate classification. Superior Classification Scheme： Among the six tested schemes， the feature selection approach based on the J-M distance algorithm， which integrates multiple feature variables， demonstrates the best performance in tobacco extraction. This scheme achieves user accuracy and producer accuracy rates exceeding 90%， with an overall accuracy of 94.88% and a Kappa coefficient of 0.94.
In order to timely， accurately and efficiently monitor tobacco Leaf Area Index （LAI）， physiological and biochemical parameters such as ground measured LAI during the whole growth period of tobacco and corresponding multi-temporal Sentinel-2 multi-spectral data were obtained. LAI inversion model based on vegetation index， LAI inversion model based on machine learning and multi-spectral data， LAI inversion model based on PROSAIL model and LAI inversion model based on multi-model coupling were respectively constructed. In addition， the accuracy of the prediction results was evaluated based on the measured LAI data， and the optimal universal prediction model for tobacco LAI during the whole growth period was selected from the perspective of single growth period and whole growth period. The results showed that， unlike the traditional LAI inversion studies in which the effects of empirical models under plot scale were mostly superior to the mechanism models， in this study， coupled with the correlation between the sensitive parameters of leaf equivalent water thickness and dry matter content and Gradient Boosting Regression Tree （GBRT）， the PROSAIL model achieved the best results in LAI inversion of tobacco during the whole growth period and each growth period， with R2， RMSE and MAE of 0.805， 0.378 and 0.276 respectively. The R2， RMSE and MAE of LAI inversion in growing stage were 0.789， 0.324 and 0.250 respectively. The R2， RMSE and MAE of LAI inversion in harvest stage were 0.576， 1.641 and 1.608 respectively.
For remote sensing data， the image segmentation model trained on the public data has scale and feature differences between the task data and the public data， which makes the traditional transfer learning ineffective in improving the model's generalization performance. To this end， an end-to-end unsupervised domain adaptation segmentation model for remote sensing image is proposed. First， residual connection， scale consistency modules and perceptual loss with class balance weights are added to the generative adversarial-based style transfer network. These reduce the image style and resolution differences between the two domains， maintain the original structural information while transferring. Second， the Visual Attention Network （VAN） that considers both spatial and channel attention is used as the feature extraction backbone network to improve the feature extraction capability. Finally， a small number of target domain samples with labels are used to fine-tune the segmentation model to further improve the segmentation accuracy. The experimental results show that the proposed model effectively alleviates the performance degradation problem caused by different features and resolutions， which is more advantageous compared with advanced domain adaptation segmentation methods.
The previous methods for hyperspectral data classification only focus on the extraction of spectral features， which often wastes some valuable spectral spatial information and leads to unsatisfactory classification results. In view of this， this paper proposes a method that combines Principal Component Analysis （PCA）， guided filtering and deep learning architecture into hyperspectral data classification. First， PCA， as a mature dimension reduction architecture， can effectively reduce the redundancy of hyperspectral information； Then， guided filtering is used to provide a simple and effective channel to obtain spatial dominant information； Finally， the stack Autoencoder model is used as a deep learning architecture to effectively process deep level multi feature image data. Train and test the algorithm using two common datasets， and then use a third common dataset to test the models trained on the other two datasets. The experimental results show that the proposed GF-FSAE algorithm achieves classification accuracy above 99%， demonstrating good classification performance and generalization ability. Compared to the CNN-AL model， the algorithm’s accuracy is slightly higher， verifying the superiority of the spectrum-spatial hyperspectral image classification framework。
In response to the challenge of significantly degraded image quality and difficulties in target detection caused by heavy fog during typhoon weather， as well as the limited generalization of existing research in power scenarios， an improved algorithm that integrated dark channel defogging with YOLOv8 is proposed. The proposed algorithm employs a two-stage processing technique. Initially， color adaptive defogging is applied based on an analysis of the image's color distribution using cumulative distribution functions and the dark channel algorithm to enhance the visibility of targets within foggy environments. Subsequently， to further enhance detection accuracy for multi-scale targets in remote sensing images，A small target detection layer is introduced into the YOLOv8 network architecture. This addition facilitates deeper feature extraction for small targets while employing MPDIoU instead of CIoU to reduce computational complexity. Experimental results demonstrate that the proposed algorithm improves detection accuracy for power towers and wind turbines by 8.1% and 3.9%， respectively. These findings validate both the feasibility and effectiveness of the proposed algorithm in processing foggy remote sensing images and recognizing targets， thereby providing reliable technical support for defogging operations on such images and identifying large-scale outdoor power facilities.
The Qinba Mountain region， encompassing the transition zone between North and South China， holds significant ecological importance. Investigating the phenological changes in vegetation and their response to climate change in this region is crucial for understanding the complexity of the ecological environment in the transitional zone and reconstructing historical climate patterns. Based on MODIS MCD12Q2 data and meteorological remote sensing data， this study employs trend analysis and correlation analysis methods to explore the spatiotemporal characteristics of vegetation phenology in the Qinba Mountain region from 2001 to 2020 and its relationship with climate change. Results indicate that the start， end， and length of the growing season in the Qinba Mountain region exhibit distinct vertical zonal distribution characteristics from east to west. The start of the growing season is primarily distributed from mid-to-late March to late April （70-110 days）， while the end of the growing season is concentrated from late October to late November （290-320 days）， with the majority of growing seasons falling between 180 and 260 days. Over the 20-year period， the overall characteristics of phenological interannual variation in the Qinba Mountain region show an average advancement of 0.38 days per year. The end of the growing season exhibits an average delay of 0.43 days per year. The length of the growing season displays an average extension of 0.80 days per year. Significant trend of change. Regarding the time lag response to climate factors， the start of phenology in the Qinba Mountain region shows the highest correlation with monthly temperature and potential evapotranspiration without significant time lag effects， while precipitation exhibits a time lag of approximately 1.73 months. Altitude to some extent determines the response relationship between the start of phenology and various meteorological elements in the Qinba Mountain region.
Understanding evapotranspiration characteristics of sandy plantation is of great significance for improving the ecological construction of sandy plantation. Using the Two-Source Energy Balance （TSEB） model by Landsat8 and the meteorological observation data as the driving data， the spatial-temporal variation characteristics of evapotranspiration of Yinkensha forest field in Hobq Desert was obtained. The critical variables of the model and evapotranspiration results were evaluated with the relevant data of Bowen ratio observation system. The results show that： （1） In the two critical variables， the net radiation estimated by the TSEB model is higher than the observed value， MAE， RMSE and MBE are 15.98， 20.02 and 0.19 W·m-2， respectively， and the correlation coefficient between the two is 0.91； The soil heat flux estimated by the TSEB model is lower than the observed value， with MAE， RMSE and MBE being 7.43， 9.29 and 2.12 W·m-2， respectively， and the correlation coefficient between the two is 0.72. （2） The evapotranspiration estimated by the TSEB model is higher than the observed value， with MAE， RMSE and MBE being 1.89， 2.02 and 3.45 mm， respectively， and the correlation coefficient between the two is 0.77. The area with large evapotranspiration is concentrated in the central and southern part of the forest farm， while the evapotranspiration in the northern part of the forest farm is relatively small. （3） Evapotranspiration estimated by the TSEB model is equivalent to that estimated by the SEBAL model， but in areas with concentrated vegetation， evapotranspiration estimated by the SEBAL model is higher than that estimated by TSEB model， and the evapotranspiration correlation of SEBAL model inversion is lower than that of TSEB model inversion. The research results can provide reference for further improving the performance of the model.
The extraction of groundwater for winter wheat irrigation in the Hebei Plain has resulted in a continuous decline in the regional groundwater levels， triggering a range of ecological and environmental issues and intensifying the contradiction between water resources and food security. Employing remote sensing technology to accurately extract the irrigation information of winter wheat is significantly important for realizing the sustainable development of water resources and rational allocation of water resources. In this study， the MOD09Q1 and TRIMS LST data were utilized to calculate the NDVI and TVDI of the Hebei Plain from October 2020 to June 2021. Based on the NDVI time series characteristics of winter wheat， the planting area of winter wheat in Hebei Plain in 2021 was extracted. Leveraging the response characteristics of TVDI to winter wheat irrigation， a method for monitoring irrigation information was determined， and the irrigated area during the different growth periods of winter wheat was inverted， subsequently obtaining the spatial distribution of the number of irrigations throughout the entire growth period of winter wheat. The results showed that the planted area of winter wheat in Hebei Plain was 20.91×103 km2 in 2021， and the extraction results were highly accurate. The less precipitation during the growth period of winter wheat had minimal impact on the TVDI of the wheat fields， while the irrigation resulted in minimum in the TVDI time series. In 2021， the irrigated area of winter wheat in the Hebei Plain with overwintering water， rejuvenation water， pulling water and grouting water was 10.23×103 km2，14.31×103 km2，9.79×103 km2， and 7.14×103 km2， respectively. The proportions of the areas with irrigation events of 0， 1， 2， 3， and 4 times for winter wheat in the Hebei Plain were 6%， 25%， 39%， 25%， and 5%， respectively. This study proposed a method to obtain the actual irrigated area of winter wheat， which provided technical support for guiding irrigation production in groundwater overexploitation areas.
Vegetation Ecological Water Demand （EWD） in arid agricultural cities plays a critical role in regional water resource management and ecosystem stability. Taking Zhangye， a typical agricultural city in the arid region of Northwest China， as a case study， this research validated remote sensing evapotranspiration products from MODIS and the PML model using flux tower observations， aiming to determine the optimal model for regional EWD assessment. Based on the validated model， a linear regression trend analysis was employed to investigate the spatiotemporal patterns of vegetation EWD from 2001 to 2023. Furthermore， a multiple linear regression model was applied to quantitatively evaluate the relative contributions of climatic variables， including air temperature （Ta）， precipitation （Pre）， and radiation （Rn）， as well as vegetation dynamics indicated by NDVI， to EWD variations. Results indicated that the PML model exhibited significantly higher accuracy and reliably represented regional variations in vegetation EWD. Between 2001 and 2023， EWD in Zhangye exhibited a significant upward trend， displaying a distinct spatial gradient increasing from north to south. The variations in EWD were jointly driven by climatic changes and vegetation dynamics. Climatic factors accounted for 59.63% of these changes （precipitation alone contributing 21.69%）， whereas vegetation changes （NDVI） contributed 43.44%， representing the primary driver for increasing EWD. This study reveals the spatial-temporal patterns and main driving factors of vegetation EWD in typical agricultural cities of Northwest China's arid regions， providing important references for efficient water resource utilization and sustainable ecological management in such areas.
Soil erosion evaluation based on soil erosion estimation model requires high-quality input data， including land use classification data， however， the traditional land use classification method has the problems of low efficiency and the phenomenon of “heterogeneity and homogeneity” when facing the task of multi-class classification. Therefore， this study tries to apply deep learning to soil erosion evaluation， adopts the semantic segmentation model with Swin Transformer as the backbone network to realize high-precision land use classification， and applies the classification results to the RUSLE model to evaluate the degree of soil erosion. It is verified that the overall accuracy of the semantic segmentation model reaches 95% and has good generalization performance. The results of soil erosion evaluation show that soil erosion in Poyang County is dominated by light erosion， and spatially presents the distribution characteristics of band and point. The results show that land use classification based on deep learning has a wider application prospect in the field of soil erosion evaluation.
The task of transmission tower detection is faced with the problem of complex target characteristics， small sample data and unbalanced feature distribution. In view of the feature difference of transmission tower under different imaging conditions， a data augmentation method based on imaging conditions is proposed in this paper. The method realizes the data augmentation of transmission tower target from two aspects of imaging angle and imaging background. The imaging angle augmentation method can generate transmission tower images at different shooting angles by simulating satellite angle changes. The imaging background augmentation method generates target images under different backgrounds by adaptively merge the method of edge inpainting and Poisson fusion. Experimental validation on Faster R-CNN， RetinaNet， and YOLOv3 frameworks demonstrates that our method achieves significant improvements in detection accuracy compared with conventional approaches. The results indicate that the proposed methodology outperforms traditional data augmentation techniques in enhancing both the detection capability and localization precision for transmission tower targets.
Accurate farmland information is of great significance for ensuring national food security. Traditional farmland extraction methods have limitations when dealing with increasingly complex remote sensing images. The development of deep learning has brought new approaches for farmland extraction. However， the classic DeepLabV3+ network has problems such as a large number of training parameters， less than ideal image segmentation accuracy， and poor generalization ability in practical applications. To address these issues， this study proposes a lightweight DeepLabV3+ network integrated with an attention mechanism. In this network， the Xception， which is the main structure of DeepLabV3+， is replaced with MobileNetV2 to reduce the number of training parameters. The channel attention mechanism SENet is introduced to improve the segmentation accuracy. The loss function is replaced with the Hybrid function to enhance the generalization ability of the model.The research results show that： （1） After lightweighting， the average training time of the network is reduced from 7.45 minutes to 1.86 minutes， and the model training parameters are decreased from 208.7 MB to 24.73 MB； （2） The precision， recall， and Mean Intersection over Union （MIoU） of the model reach 90.26%， 90.08%， and 81.58% respectively， all of which are superior to those of other comparative models； （3） The absolute value of the relative error of the extraction results in Gaomiao Village， Yongchuan District， is 7.95%， which performs better than the DeepLabV3+ network model. In conclusion， the improved DeepLabV3+ network can effectively improve the farmland extraction effect. While reducing the number of parameters and increasing the operation speed， it also improves the extraction accuracy. It has certain universality and transferability and can provide accurate spatial distribution information of farmland for agricultural planning.
Above Ground Biomass （AGB） is an important indicator of grassland ecosystem function and grassland productivity. Accurate estimation of AGB is of great significance for grassland management and ecological environment assessment. Taking part of grassland in Xilinhot， Hulunbuir and Ordos grassland of three different types as the research area， based on UAV multi-spectral data， LiDAR data and field measured sample data， multiple texture and vegetation index were obtained， and different feature combinations were obtained through various feature screening methods. Five regression analysis algorithms， including Random Forest（RF）， Multiple Linear Regression（MLR）， BP neural network（BP）， Support Vector Regression（SVR）， and Long Short-Term Memory neural network（LSTM）， were used to construct a grassland AGB estimation model， and the optimal estimation model was obtained after comparison and evaluation， and the spatial biomass estimation was carried out. The results indicate that： （1） The feature combination selected by the feature importance method， including spectral indices and the measured average plant height （Mean Height） within the sample plots， achieved high accuracy in multi-model comparisons for grassland AGB estimation； （2） The Random Forest model outperformed other models in estimation accuracy. Using the coefficient of determination（R²），Root Mean Squared Error（RMSE），and Mean Absolute Error（MAE） as evaluation metrics，the test sample achieved an R² of 0.859， with RMSE and MAE values of 35.17 g/m² and 28.22 g/m²， respectively. The study demonstrates that integrating multi-source remote sensing features with machine learning algorithms can effectively overcome the limitations of traditional AGB estimation methods. The Random Forest model based on optimized feature combinations provides a reliable methodological reference for accurate grassland AGB estimation.
Geostationary meteorological satellite imagers enable extensive and continuous monitoring of sea fog. However， when cloud layers cover the fog region， the signals received by the satellite primarily originate from the upper cloud layers， making it challenging to ascertain the presence of sea fog in the lower layers. In the East China Sea， occurrences of fog events beneath cloud cover are frequent， impeding the operational application of satellite remote sensing in sea fog detection. Based on the channel design features of the Advanced Himawari Imagers （AHI） on-board the Himawari-8 geostationary meteorological satellite， we creatively proposed an algorithm aiming at detecting sea fog covered by high-level ice clouds combined with radiation transfer theory simulation and real observations. Two long-wave infrared channels （8.5 μm and 11 μm） were utilized to identify the cloud-top phase （ice clouds or water clouds）. Low-level water clouds and sea fog beneath ice clouds were distinguished by the differences in radiation characteristics in the short-wave infrared （1.6 μm and 2.25 μm） and visible （0.64 μm） channels when ice clouds were identified. Finally， the accuracy of this algorithm was verified according to Cloud-Aerosol Lidar with Orthogonal Polarization （CALIOP） data. The results showed good consistency between our algorithm and CALIOP， with an average probability of detection of 49%， a false alarm ratio of 7%， and a critical success index of 46%. Due to the difficulty of detecting fog under clouds in the field of satellite remote sensing， the results demonstrate that this algorithm can detect sea fog under ice clouds.
Regional poverty is the core manifestation of the imbalance in regional social and economic development. Rapid， objective and accurate acquisition of regional poverty measurement data is of vital importance for revealing the evolution process of regional poverty and formulating targeted poverty alleviation decisions. This study takes 117 counties and districts in Shanxi Province， an important battlefield of poverty alleviation， as the research object. Based on the night light data and statistical data from 2011 to 2020， with the help of the vulnerability-sustainable livelihood framework， a regional optimal multi-dimensional poverty measurement correlation model is constructed. And with the help of spatial self-correlation and cold and hot spot analysis， the poverty degree and spatio-temporal evolution process are revealed. The research results show that： （1） The constructed cubic polynomial multi-dimensional poverty correlation model has high accuracy， with the accuracy rate and average relative error ranging from 81.03% to 84.48% and 18.01% to 22.63% respectively； （2） In terms of time： The degree of multi-dimensional poverty in Shanxi Province has generally shown an improving trend. The number of counties and districts with extremely low and extremely high MPIe grades has decreased significantly， while the number of counties and districts with lower， medium and higher grades has shown a wavy upward trend. The polarization of poverty degrees is narrowing； （3） Spatially： The high values of MPIe grid spatialization occur in urban areas， mainly concentrated in areas with relatively flat terrain. The low values of MPIe are mainly distributed in the Luliang Mountain area and the Yanshan-Taihang Mountain Area， and have the characteristics of continuous distribution. Moran's I index reveals that regional poverty shows an agglomeration trend， and the degree of poverty in southern Shanxi is better than that in northern Shanxi， and in the east it is better than that in the west. The degree of poverty is alleviating and the polarization of poverty is narrowing. The LISA index reveals that in Shanxi Province， the low-concentration distribution of multi-dimensional poverty is the largest， while the high-concentration distribution is relatively small. The analysis of cold and hot spots reveals that the hot spots present the characteristics of contiguous concentration， corresponding to the high-concentration areas. The secondary hot spots are distributed around the hot spots， while the number of cold and secondary cold spots is very small. It effectively proves the rationality and objectivity of the night light data for poverty simulation， and the multi-dimensional poverty model can better represent Shanxi Province.
