Examining the influence of bare soil UAV imagery combined with auxiliary datasets to estimate and map soil organic carbon distribution in an erosion-prone agricultural field

Abstract

Soil organic content (SOC), an indicator of soil fertility, can be estimated quickly and accurately with remote sensing (RS) datasets; however, the issue of vegetation cover on the field still remains a major concern. In order to minimize the effects of vegetation cover, studies relating reflectance spectra to SOC may require bare soil. However, acquiring satellite images devoid of vegetation is still an enormous challenge for RS techniques. This is because the area that may have been accurately predicted at a targeted date is sometimes limited since many pixels are covered by vegetation. The study goal was to assess the impact of using UAV-borne imagery coupled with auxiliary datasets, which include spectral indices (SPIs) and terrain attributes (TAs) (at 20 cm and 30 m resolution), singly or merged, to estimate and map SOC in an erosion-prone agricultural field. Both field samples and UAV imagery were acquired while the fields were bare. Using a grid sampling design, 133 soil surface samples were collected. The models used include partial least square regression (PLSR), extreme gradient boosting (EGB), multivariate adaptive regression splines (MARS), and regularised random forest (RFF). The models were evaluated using the root mean squared error (RMSE), the coefficient of determination (R2), ratio of performance to interquartile distance (RPIQ), and the mean absolute error (MAE). For prediction, the three merged datasets (R²val = 0.86, RMSEval = 0.13, MAEval = 0.11, RPIQval = 4.19) outperformed the best separate dataset (R²val = 0.82, RMSEval = 0.15, MAEval = 0.10, RPIQval = 2.08). Though all datasets detected both low and high estimates of soil SOC, the three merged datasets with EGB showed a less extreme prediction error. This study demonstrated that SOC can be estimated with high accuracy using completely bare soil UAV imagery with other auxiliary data, and it is thus highly recommended.

Keywords: Agricultural soil; Machine learning algorithms; Soil organic carbon; Spectral indices; Terrain attributes; UAV.