Soil Organic Carbon Prediction Based on Different Combinations of Hyperspectral Feature Selection and Regression Algorithms

Abstract

Cropland soil organic carbon (SOC) is crucial for global food security and mitigating the greenhouse effect. Accurate SOC prediction using hyperspectral data is essential for dynamic monitoring of soil carbon pools in croplands. However, effective methods to reduce hyperspectral data dimensionality and integrate it with suitable regression algorithms for reliable prediction models are poorly understood. In this study, we analyzed 108 soil samples from Changting County, Fujian Province, China. Our objective was to evaluate the performance of various combinations of six feature selection methods and four regression algorithms for SOC prediction. Our findings are as follows: the combination of the Successive Projections Algorithm (SPA) and Partial Least Squares (PLS) yielded the most favorable results, with R2 (0.61), RMSE (1.77 g/kg), and MAE (1.48 g/kg). Moreover, we determined the relative importance of variables, with the following ranking: 696 nm > 892 nm > 783 nm > 1641 nm > 1436 nm > 396 nm > 392 nm > 2239 nm > 2129 nm. Notably, 696 nm exhibited the highest importance in the SPA-PLS model, with the Variable Importance in Projection (VIP) value of 1.22. This study provides profound insights into feature selection methods and regression algorithms for SOC prediction, highlighting the superiority of SPA-PLS as the optimal combination.

Keywords

• soil organic carbon
• hyperspectral
• feature selection
• regression
• relative importance