Mapping stocks of soil total nitrogen using remote sensing data: A comparison of random forest models with different predictors

Estimation of soil total nitrogen (STN) is important for quantifying the spatial distribution of soil N (Nitrogen) stocks, determining the amount absorbed by crops and avoiding yield losses and environmental pollution. The objective of this research was to determine the spatial distribution of STN using a remote sensing approach and to investigate the importance of different predictors under two conditions: the bare soil condition and the vegetated condition. We also investigated whether the Sentinel-2A red-edge bands and relative spectral indices were suitable for STN estimation. Soil data were collected from the topsoil (0-20 cm) at 104 sampling sites on 06/15/2016 from farmland in a black soil region in Northeast China. Two Sentinel-2A Multispectral Instrument (MSI) remote sensing images were acquired on 05/26/2016 and 08/03/2016 to represent the two conditions. Environmental variables included the terrain attributes, temperature and precipitation. Then, 21 predictors, including the original bands (O), normal spectral indices (S), red-edge indices (R) and environmental variables (E), were employed to estimate the spatial distribution of the STN content using a random forest (RF) model. Finally, different predictors were combined to construct RF models, and the prediction model with the best performance was selected to determine the spatial pattern of the STN content. The results showed that the predictors had different levels of importance under the two conditions. However, most environmental variables and normal spectral indices always play a significant role in the estimation of STN. The model with the combination of the original bands, normal spectral indices, red-edge indices and environmental variables (O + S + E + R) under the bare soil condition had the best prediction performance, and the combination of the original bands, normal spectral indices and red-edge indices (O + S + R) model had a performance similar to the O + S + E + R model. Therefore, the selection of suitable predictors is necessary to predict STN. The spatial pattern of STN was related to the crop type and elevation of the study area. The results of this study suggested that the proposed RF-based remote sensing method was able to accurately capture the variation in STN and that the performance of the prediction model can be improved by providing enough types of suitable predictors.