Machine learning based digital mapping of soil properties in mid-Himalayan terrain

Abstract

Abstract Soil physico-chemical properties influence ecosystem services and subsequently human’s lives, therefore soil information is crucial for promoting sustainable land use and ensuring the long-term health and productivity of soils. In environmentally vulnerable regions like the Himalayas, where rapid socio-economic development is seen and expected to grow, it is imperative to precisely map the soil information in the landscape to protect and manage it sustainably. The demand for applying artificial intelligence to automate a variety of tasks for its ability to learn and analyze large datasets has enabled the applications of different machine learning methods for digital soil mapping (DSM) approach. Despite the growing number of ML algorithms used in DSM, no studies have used preprocessing technique like resampling for soil datasets for supervised ML regression model. The main objective of this study is the mapping and analyses of soil texture and organic carbon mapping using a random forest regression (RFR) model of an area in the mid-Himalayas by employing more than 100 environmental covariates. The study uses gaussian noise up-sampling technique to resample the small imbalanced soil datasets from the highly undulating terrain, resulting in significantly accurate maps. Model performances, evaluated against an unknown dataset were significant with an R-square of 0.80, 0.79, 0.72, and 0.84 for clay, sand, silt, and SOC, respectively, and their respective mean absolute error and root mean square error are reported. Further, sensitivity analysis of the environmental covariates contributing to the model resulted in effective contribution of all the soil forming factors.