Comparing laboratory and airborne hyperspectral data for the estimation and mapping of topsoil organic carbon: Feature selection coupled with random forest

Abstract

Rapid monitoring of soil organic carbon (SOC) with fine sampling resolution is vital for further understanding of the global carbon cycle and sustainable management of soil resources. Proximal visible and near-infrared (Vis–NIR) spectroscopy is an effective approach to analyze SOC. However, this technique can only be used for point-to-point monitoring and not for grid pixels evenly spread throughout the area. Airborne hyperspectral imagery with high-spectral- and spatial-resolution provides a promising tool for mapping topsoil SOC at a fine scale, but suffers from the interference of some external factors. Using 45 topsoil samples collected from an agricultural field in the United States, this study aimed to compare the potential of airborne hyperspectral image in estimating and mapping of bare topsoil SOC with that derived from proximal laboratory Vis–NIR spectral data. Random forest (RF) along with two advanced feature selection algorithms, namely, continuous wavelet transform (CWT) and competitive adaptive reweighted sampling (CARS), was applied to optimize the performance of the prediction models. Results showed that laboratory and airborne spectra presented similar spectral shapes and strengths, but laboratory spectral curves were smoother than airborne spectral curves, which were noisier. Laboratory spectra (R2 = 0.79–0.87) performed better than airborne hyperspectral imagery (R2 = 0.49–0.76) in cross-validation, regardless of feature selection algorithms. The CWT-RF models resulted in the highest cross-validation results for laboratory (R2 = 0.87) and airborne (R2 = 0.76) spectra, suggesting their robustness in SOC prediction. The SOC maps retrieved from full-spectrum-RF, CWT-RF, and CARS-RF models all exhibited similar spatial distribution patterns. With airborne hyperspectral imagery serving as a valuable data source at pixel level for digital soil mapping, the methodological framework proposed in this paper could improve the accuracy and reduce the prediction uncertainty of SOC maps by selecting and adopting the optimal subset of spectral variables.