Modelling of soil depth and hydraulic properties at regional level using environmental covariates- A case study in India

Soil hydraulic properties are important for agroecosystem modelling, irrigation scheduling, drought risk assessment, and preparation of proper land use planning. Field capacity (FC) and permanent wilting point (PWP) are the two vital soil hydraulic properties that determine the availability of water for plant growth. In the present study, we evaluated four data mining algorithms for mapping of field capacity and permanent wilting point along with soil depth over 16.2 M ha of Andhra Pradesh state, South India. Class pedo-transfer function was developed based on the soil texture classes (760 observations) and the profile total field capacity (TFC) and total profile permanent wilting point (TPWP) were computed for 1 m soil depth. Multi-linear Regression (MLR), Cubist, Support Vector Machine (SVM), and Random Forest (RF) models were evaluated using 2267 soil profile datasets collected over the study area. Along with Landsat-8 data and climatic datasets, terrain attributes such as plan curvature, profile curvature, topographic wetness index (TWI), topographic position index (TPI), Multiresolution Index of Valley Bottom Flatness (MrVBF), and Multi-resolution Ridge Top Flatness (MrRTF) were used as environmental covariates. The models were calibrated using 80% of total field observations and validated using 20% of observations. The validation results showed that RF outperformed all other models for the prediction of both soil depth and profile TFC and TPWP. RF explained 39% of the variation of total field capacity and permanent wilting point with RMSE of 133 mm and 89 mm, respectively. The highest mean TFC and TPWP were found in Vertisols (256 and 154 mm, respectively) and Inceptisols (216 and 126, mm respectively). The highresolution (250 m) maps of hydraulic properties and soil depth are useful for land use planning and different crop modelling studies.

Automated summary

This study evaluated four data mining algorithms for mapping soil field capacity and permanent wilting point along soil depth in Andhra Pradesh, South India. The Random Forest model outperformed others, explaining 39% of variation and achieving the best predictive results for these soil hydraulic properties.