A Comparative Study on the Predictive Ability of Machine Learning Techniques for Spatial Mapping of Soil Properties in Indian Himalayan Region

Digital soil mapping (DSM) has emerged as a powerful tool for predicting soil properties and classes across large geographical areas, aiding in sustainable land management and environmental planning. In this study, we explore the application of various machine learning (ML) models and compared their predictive ability for digital mapping of soil properties in a hilly watershed located in the Indian Himalayan region in Dehradun district, Uttarakhand, India. The study area covers approximately 286.03 hectares and is characterized by diverse soil properties and complex terrain. Surface soil samples collected from 116 georeferenced locations using a grid sampling approach formed the soil database used in the current study. Various remote sensing data derived environmental covariates like spectral indices and terrain factors were used by different ML techniques as inputs to make accurate and spatially continuous predictions of different soil properties. Multiple machine learning algorithms, including multivariate adaptive regression splines (MARS), support vector regression (SVR), and artificial neural network (ANN) were employed for predicting soil properties and their performances were compared with multiple linear regression (MLR) models. Results demonstrated the superior predictive performance of ML algorithms over MLR, highlighting the efficacy of nonlinear modeling in capturing intricate relationships within the field of digital soil mapping. MARS as feature selection technique improved model performance and interpretability. MARS-SVR and MARS-ANN had low RMSE, MAD, and MAPE values when predicting soil characteristics. The study yielded promising outcomes, demonstrating the effectiveness of machine learning models in digital soil mapping. The models produced high-resolution soil maps, enabling better land use planning, precision agriculture, and natural resource conservation. The ML algorithms were additionally utilized to generate prediction maps of soil characteristics that might serve as the basis for various soil investigations and initiatives.