Estimation of Soil Moisture with Meteorological Variables in Supervised Machine Learning Models

Abstract

Water cycles, climate-related hazards, and agroirrigation are strongly controlled by soil moisture (SM) content. For water resource management, prediction is a key to mitigate and regulate expected economic losses and property damages. This paper compares two supervised machine learning (ML) techniques: support vector regression (SVR) and random forest (RF) to predict SM. In RStudio, various meteorological variables: temperature, relative humidity, wind speed, and rainfall are trained to estimate SM. For eight divisions, SM and weather variables are obtained from the National Aeronautics and Space Administration (NASA) Prediction of Worldwide Energy Resources (POWER). The experiments include daily observations for 39 (1982 to 2020) to develop SVR and RF models. To estimate SM from the predictive model, datasets from diverse regions: Rajshahi, Mymensingh, Chittagong, and Sylhet are utilized in training (60%) and Rangpur, Barisal, Khulna, and Dhaka are segregated for validation (40%) resulting in accuracy of 88 to 95.8%. This model further is applied to forecast daily SM for each city including two districts (Bogra and Jessore) and found slightly higher model performance for SVR (90.7%) than RF (90.1%) on average (Year: 2021). For agricultural, industrial and urban water supplies as well as drought, landslides, and river erosions can be mitigated by an accurate estimation of soil moisture. The investigations encourage for providing SM budget to public with supervised ML techniques mostly among data-sparse regions.