Numerical Estimation of Surface Soil Moisture by Machine Learning Algorithms in Different Climatic Types

The present study was designed to provide a model for surface soil moisture numerical estimation. This assessment is done based on the direct ground measurement of soil moisture in 5 cm (SM5) and 10 cm (SM10) depths using machine learning models. To do this, various meteorological variables (16 variables) were used as model inputs. The data were evaluated on a daily scale during 2017–2020. Of these data, 75% of days were randomly considered as train and 25% as test. The components relevant to air and soil temperature, relative air humidity, evaporation, and vapor pressure are the most important factors that affect daily soil moisture. A mixture of these variables is used as model input. For this purpose, two machine learning models, including a multilayer perceptron (MLP) neural network and an adaptive neuro-fuzzy inference system (ANFIS) were used. Three agriculture meteoritical stations located in three different climates were assessed: (1) Gharakhil Station (semi-humid and moderate), Zarghan Station (semi-arid and cold), and Zahak Station (extra-arid and moderate). According to the comparison between estimates and measurements, both models had a relatively desired performance in Gharakhil and Zarghan (57% < R2 < 66% for SM5 and 45% < R2 < 58% for SM10). However, the performances were weak and almost unacceptable in the extra-arid Zahak climate (14% < R2 < 17% for SM5 and 18% < R2 < 22% for SM10). According to the relative root mean square error (RRMSE) and Nash–Sutcliffe value of stations, the models in humid climates, performed better than those in arid and extra-arid climates. The best RRMSE value was obtained by ANFIS in Gharakhil Stations (0.193 for SM5 and 0.178 for SM10), while the weakest RRMSE value was obtained in Zahak Station, which equaled 0.887 (via MLP) and 0.767 (via ANFIS) for SM5 and SM10, respectively. The applied models were not superior to each other; however, the ANFIS model was slightly superior to MLP in most cases.