Spatial prediction of major soil properties using Random Forest techniques - A case study in semi-arid tropics of South India.

Abstract

The purpose of the study is to map the spatial variation of major soil properties in Bukkarayasamudrum mandal of Anantapur district, India using Random Forest model. The study area is divided into different Physiographic Land Units (PLU) based on landform, landuse and slope. Random Forest model (RFM) was developed based on field survey data of 116 surface samples (0–30 cm) representing all major PLU units of the study area. RFM is neither sensitive to over fitting nor to noise features and has capacity to handle large datasets. High resolution satellite imagery (IRS LISS IV data- 3 bands), terrain attributes such as elevation, slope, aspect, topographic wetness index, topographic position index, plan & profile curvature, Multi-resolution index of valley bottom flatness and Multi-resolution ridge top flatness, Vegetation factors like NDVI, EVI and land use land cover (LULC) are used as covariates along with legacy soil data of 1:50,000 scale. The predicted organic carbon, pH and EC ranged from 0.24–1.03%, 6.9–9.0, 0.11–0.97 dsm− 1 respectively. The model performance was evaluated based on Coefficient of determination (R2) and Lin's Concordance coefficient (CCC). The model performed well with R2 and CCC values of 0.23 and 0.38 for SOC, 0.30 and 0.37 for pH, and 0.62 and 0.70 for EC respectively. Variable importance ranking of RFM model showed that EVI and NDVI are the most important predictors for organic carbon whereas drainage and NDVI for EC and pH respectively. This technique can be applied to similar landscapes with more observations to refine the spatial resolution of soil properties.