Development of a linear based stochastic model for daily soil temperature prediction: One step forward to sustainable agriculture

Abstract Many variables in our environment are impacted by alternations in soil temperature (Tsoc). Fluctuations in Tsoc alter gas absorption and emission capacity of the soil, which mutually influence climate changes. Hence, providing viable methodologies of modeling Tsoc is of great importance. Since thorough investigations of time series structure are mostly neglected, and appropriate pre-processing methods are not applied to them, the direct use of nonlinear methods for soil temperature forecasting has become more common than other approaches. In this study, unlike most of the existing studies that estimate Tsoc as a variables based approach, soil temperature is forecasted using a linear stochastic based methodology with sufficient knowledge of time series structure. With this methodology, the components of the Tsoc time series are determined in order to perform stochastic modeling using Holt-Winters advanced exponential smoothing. The results of this procedure are compared with two stochastic techniques based on seasonal standardization (stdω) and spectral analysis (sf) in terms of six Tsoc time series at each station. The Tsoc. time series applied in the current study were measured at Bandar Abbas and Kerman synoptic stations, Iran, at depths of 5, 10, 20, 30, 50, and 100 cm. Owing to the comparison of the applied techniques in different Tsoc time series analysis, the stochastic model with Holt-Winters advanced exponential smoothing (Bandar Abbas station: (R2% = 96.649, RMSE = 0.977, MAPE% = 1.963, AICc = −733.006; Kerman station: R2% = 92.416, RMSE = 1.806, MAPE% = 7.793, AICc = 360.410) outperformed the stdω and sf methods. In addition to comparing linear stochastic methods, the results of Tsoc modeling are compared with two powerful nonlinear methods. In an antecedent study on the mentioned sites, Nahvi et al. (2016) employed extreme learning machine (ELM) and self-adaptive evolutionary ELM (SaE-ELM) nonlinear techniques. A comparison of results indicates that the proposed methodology outperformed the nonlinear models (ELM and SaE-ELM). Indeed, the proposed linear based stochastic model not only enhanced forecasting accuracy but also demonstrated the capability of linear models in Tsoc time series forecasting. Concerning the comprehensibility and generality of the framework, it can apply to other stations with different climatological conditions.