高光谱遥感地表元素反演方法研究及地质应用前景OA

Hyperspectral remote sensing technology,with its continuous spectral information,provides an important data basis for the rapid quantitative inversion of surface elements.However,under different land cover conditions,the relationship between element concentrations and spectral features often exhibits complex nonlinearity,which directly restricts the accuracy and generalization capability of inversion models.This paper focused on Songnen Plain,which serves as a critical area for land,mineral,and ecological resources in northern China,and the distribution patterns of its surface key geochemical elements are of great significance for regional ecological environment research.To achieve high-precision quantitative inversion of geochemical elements,this study focused on soils under different land cover types within the Nenjiang river basin and Hulan River basin.By integrating hyperspectral remote sensing data with geochemical analysis data,we systematically investigated the spatial heterogeneity and spectral response mechanisms of nitrogen(N),phosphorus(P),and potassium(K).Nine spectral preprocessing techniques,including first derivative,natural logarithm,continuum removal,and multiplicative scatter correction,were initially employed to identify characteristic bands strongly correlated with element contents.Optimal spectral responses were observed for N within the 510～570 nm range;P exhibited significant correlations at 548 576,1 832 nm,and 2 174 nm;while K demonstrated high sensitivity at 372 734,and 1 912 nm.The first derivative normalization transformation showed superior performance and was subsequently utilized as input variables to construct both traditional multiple linear regression(MLR)models and convolutional neural network(CNN)-based deep learning models.The results demonstrated that the CNN model,leveraging its deep feature extraction capabilities,significantly outperformed MLR in handling complex nonlinear relationships and data heterogeneity within soil spectra,achieving test set R²values of 0.994,0.823,and 0.913 for N,P,and K prediction across diverse land covers,respectively.This confirms the superiority and stability of deep learning for geochemical element inversion under heterogeneous surface conditions.Furthermore,the inversion models were synergistically applied to airborne hyperspectral data with 3.75-meter spatial resolution.Combined with inverse distance weighting interpolation,this approach generated high-precision spatial distribution maps that clearly delineated regional-scale element gradients and micro-scale features at the field scale.The proposed hyperspectral air-ground collaborative inversion framework provides a transferable technical paradigm for rapid and high-precision detection of surface geochemical elements.The deep learning collaborative inversion framework developed in this study can potentially be extended to investigate geochemical fields of uranium and associated elements in the future,offering new technical means for uranium geological exploration.