Improvement of spatial prediction of soil depth via earth observation

Abstract

Soil depth is one of the most critical factors which impact on culture productivity and makes difficult appropriate management decisions. However, assessing this parameter is also the most challenging tasks in the agronomic field. The objective of this work was to predict the spatial distribution of soil depth from space techniques (remote sensing, RS) and machine learning. A total of 292 sites were allocated (based on the toposequence approach) and drilled (0–2 m depth) at three different locations in Brazil. Based on these, in-situ traditional depth maps (denominated field-map) were elaborated for validation. Afterwards, we created a strategy to achieve these different depths by RS (RS) approach. Landsat 8 OLI bands, Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI) and emissivity in dry and rainy seasons as well as terrain attributes were applied to predict soil depth. For this purpose, the most important covariates were selected using Recursive Feature Selection (RFE) based on Random Forest (RF) and Support Vector Machine (SVM). Afterwards, the application of RF and SVM by selected covariates were compared based on tenfold cross validation for each location. The best model was selected based on R2, RMSE and MAE, accuracy and bootstrapping approach and uncertainties. Terrain attributes were important to discriminate soil depth. Although, LST and NDVI also presented important contribution to this task. Different seasons implies on water and plant dynamics in deep and shallow soils. This impacted on NDVI and LST as detected by RS. Thus, the method brings more variables to infer soil depth. The RF model performed better than SVM to predict soil depth with an average of 0.77 R2. The accuracy between a digital soil mapping and a field-map reached 0.58 to 0.81 indicating an important result considering the difficulty of the objective This may help pedologists and farmers as well as water and plants environmental monitoring.