Comparative Assessment of Different Machine Learning Models to Estimate Daily Soil Moisture

Abstract

Soil moisture is vital as it is the primary governing factor of agriculture production and natural vegetation growth. It plays an essential role in understanding the hydrological cycle and its effect on weather and climate, and its precise prediction helps to manage the water resources optimally. Prediction of soil moisture is dependent on surface meteorological variables and soil attributes. Existing soil moisture models/prediction methods are inaccurate, and developing an optimum mathematical model for it is difficult. This study evaluates the performance of four machine learning models (deep neural network (DNN) regression, support vector machine (SVM), multiple layer perceptron (MLP), and multi-linear regression (MLR) to estimate the soil moisture conditions. The models were tested for soil moisture at two depths (25 and 50 cm depth) using the meteorological data of two stations located in a Lesser Himalayan catchment. The model outputs were compared with the observed data, and intercomparison was also made. The model performance was evaluated based on MAPE, RMSE, Nash–Sutcliffe efficiency coefficient (E<inf>N–S</inf>), and R2. The study results indicated that the DNN model outperforms the other prediction models with the highest efficacy for both stations. Therefore, the DNN model can be endorsed to estimate soil moisture when primary meteorological data are available, and it can be promising for water-efficient agriculture applications and draught management. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.