基于GF-3雷达数据极化分解与深度学习的干旱区绿洲土地覆被及盐渍化分级研究OA

Land use/land cover(LUCC)change is a key scientific issue for understanding human-land interactions,and its accurate monitoring provides essential support for regional sustainable development decision-making.This study takes the Keriya Oasis as the study area and constructs a multi-source remote sensing collaborative classification framework by integrating GaoFen-3(GF-3)fully polarimetric synthetic aperture radar(SAR)data,Landsat 8 OLI multispectral imagery,and in situ measurements of soil physicochemical properties.Soil salinization levels(slight,moderate,and severe)are employed as the core quantitative indicator of land cover quality and,together with land use types(cropland,vegetation,water bodies,and bare land),form the classification scheme.By applying eight polarimetric decomposition methods,a random forest-based feature selection algorithm,and a U-Net deep learning model,this study systematically explores optimal interpretation strategies for land use/land cover classification in arid oasis environments.The experimental results demonstrate that,compared with traditional image classification algorithms,the U-Net deep learning framework exhibits a significant advantage in classification accuracy,achieving an overall accuracy of 78.21%and a Kappa coefficient of 0.72.The model effectively integrates radar backscattering features,optical spectral information,and soil physicochemical parameters such as soil organic matter content.By constructing a multidimensional feature space,it successfully addresses the problem of spectral heterogeneity within vegetation-salinization mixed pixels.The proposed multi-source data fusion-based classification approach provides a novel technical pathway for oasis ecosystem monitoring.Moreover,the spatial heterogeneity captured by the classification results offers a robust scientific basis for land degradation control and resource management decision-making in oasis regions.