Soil temperature estimation at different depths using machine learning paradigms based on meteorological data

Abstract

Knowledge of soil temperature (ST) is important for analysing environmental conditions and climate change. Moreover, ST is a vital element of soil that impacts crop growth as well as the germination of the seeds. In this study, four machine-learning (ML) paradigms including random forest (RF), radial basis neural network (RBNN), multi-layer perceptron neural network (MLPNN), and co-active neuro-fuzzy inference system (CANFIS) were used for estimation of daily ST at different soil depths (i.e. 5 cm: ST₅; 15 cm: ST₁₅; and 30 cm: ST₃₀) during 2016–2019 at Bathinda weather station, located in South-western Punjab (India). Five different combinations were formulated using four meteorological data, namely T_mean (mean air temperature), RH (relative humidity), WS (wind speed), and SSH (bright sunshine hours), and the optimal one was nominated by employing the gamma test (GT) for each soil depths, respectively. During the validation period, the outcomes of the RF, RBNN, MLPNN, and CANFIS models were evaluated according to performance metrics such as mean absolute error (MAE), root mean square error (RMSE), scatter index (SI), coefficient of efficiency (COE), Pearson correlation coefficient (PCC), and index of agreement (IOA), as well as through pictorial interpretation (Taylor diagram, box-whisker plots, time-variation, scatter plot, and radar chart). The comparison of the results of ML paradigms revealed the highest accuracy was achieved by the CANFIS model at all depths with MAE (RMSE) = 0.788, 0.636, 0.806 (1.074, 0.854, 1.041) °C, SI = 0.040, 0.033, 0.040, and COE (PCC)/IOA = 0.986, 0.991, 0.985 (0.994, 0.995, 0.993)/0.996, 0.998, 0.996. Thus, the results highlight the capability of the CANFIS model with T_mean, RH, WS, and SSH inputs for daily ST estimation at different soil depths on the study site.