Digital mapping of soil organic carbon density using newly developed bare soil spectral indices and deep neural network

Abstract

Soil organic carbon density (SOCD) is an important parameter of agricultural soils and is useful for the improvement of environment and agricultural production. Proximal and remote sensing techniques are effective methods for digital soil mapping of SOCD. The current study used three types of spectral data, including laboratory proximal spectra, airborne hyperspectral and Sentinel 2 multispectral images, to predict SOCD in an agricultural land. Bare soil spectral indices (BSSIs) were developed to predict SOCD and compared with the published vegetation indices (VIs). With 45 soil samples, the partial least square regression (PLSR), back propagation neural network (BPNN) and deep neural network (DNN) prediction models were established to map SOCD. The results showed that the laboratory spectra (R2 = 0.70–0.80) had the best performance of SOCD prediction, followed by the airborne (R2 = 0.43–0.81) and Sentinel 2 (R2 = 0.14–0.57) spectra. The SOCD maps derived from airborne and Sentinel 2 images had similar spatial distribution trends. The BSSIs (R2 = 0.24–0.81) obtained higher accuracy than the VIs (R2 = 0.14–0.74) in SOCD prediction. Moreover, DNN model was the best for digital soil mapping of SOCD among three prediction models. This study offered an effective approach for mapping SOCD in bare soil areas.