GF-5 hyperspectral inversion of soil parameters using a VAE style-based spectral fusion model

Inverting soil parameters through hyperspectral techniques is currently one of the highly popular research topics and the major challenges in quantitative remote sensing. To date, indoor spectral data-based inversion models cannot be directly applied to satellite-based hyperspectral data, due to the weak model migration capability caused by the large differences between the two spectral data. Therefore, the present study aims to improve the inversion soil parameter accuracies using satellite-based GF-5 hyperspectral remote sensing data by merging multiple hyperspectral data. First, indoor Analytical Spectral Devices (ASD) hyperspectral and pre-processed GF-5 data of soil samples were used to develop a variational auto-encoder (VAE)-based spectral fusion model capable of transforming GF-5 spectra into indoor spectra. Second, traditional machine learning regression algorithms, namely Partial Least Squares Regression (PLSR) and Support Vector Regression (SVR), were used to build an inversion model using the mixed spectra data to determine the spatial distributions of soil organic matter (SOM), arsenic (As) and copper (Cu) contents across a large study area. The results demonstrated the effectiveness of the VAE-based spectral fusion model in removing substantial noise information while preserving the spectral features from the GF-5 data. The optimal inversion accuracies of the SOM, As, and Cu contents showed coefficients of determination (R²) of 0.87, 0.88, and 0.85, which are 38%, 55%, and 28% higher than those obtained using the original GF-5 data-derived model, respectively. In addition, the spatial distributions of the SOM, As, and Cu contents demonstrated that the GF-5 satellite data are more intuitive and effective for large-scale soil composition analysis.