Predicting soil organic matter with ZY1E hyperspectral images by correcting soil spectrum and expanding sample size

Abstract

Hyperspectral images provide an efficient means for large-scale predicting of soil organic matter (SOM) content, yet its accuracy is often hindered by soil moisture effects and limited soil sample size. To address these challenges, this study proposes a novel method that integrates soil spectrum correction and sample expansion to improve SOM content prediction accuracy. An improved orthogonal signal correction (OSC) algorithm using visible and shortwave infrared drought index (VSDI) as a reference is developed to correct soil spectra and reduce external parameter reliance. Additionally, a sample expansion algorithm is developed to enhance sample diversity and reduce overfitting, integrating the soil spectral response mechanism with the spatial autocorrelation among samples. Finally, the hybrid Back Propagation Neural Networks - Random Forest (BPNN-RF) model is applied to predict SOM content. The proposed method was validated by 80 topsoil samples and ZiYuan-1 02D (ZY1E) hyperspectral images in Nong'an County, Jilin Province, China. The results indicate that the improved OSC algorithm effectively corrected the soil moisture effects and enhanced spectral sensitivity to SOM, increasing the average absolute correlation coefficient from 0.34 to 0.41, with a maximum value exceeding 0.50. Sample expansion improved model performance (the coefficient of determination (R²) increased from 0.42 to 0.71, the root-mean-square error (RMSE) decreased from 0.34 % to 0.24 %), and combining it with soil spectral correction further raised R² to 0.81 and reduced RMSE to 0.19 %. SHAP analysis revealed that the top 20 important bands fell within SOM-sensitive ranges. The distribution pattern of predicted SOM content map was inverse to that of the Digital Elevation Model (DEM) map yet consistent with that of the annual average precipitation map. Thus, this method improves the spatiotemporal adaptability of SOM prediction using hyperspectral images, offering a robust approach for rapid and large-scale soil monitoring.