Multi-step ahead soil temperature forecasting at different depths based on meteorological data: Integrating resampling algorithms and machine learning models

Abstract

Direct soil temperature (ST) measurement is time-consuming and costly; thus, the use of simple and cost-effective machine learning (ML) tools is helpful. In this study, ML approaches, including KStar, instance-based K-nearest learning (IBK), and locally weighted learning (LWL), coupled with resampling algorithms of bagging (BA) and dagging (DA) (BA-IBK, BA-KStar, BA-LWL, DA-IBK, DA-KStar, and DA-LWL) were developed and tested for multi-step ahead (3, 6, and 9 d ahead) ST forecasting. In addition, a linear regression (LR) model was used as a benchmark to evaluate the results. A dataset was established, with daily ST time-series at 5 and 50 cm soil depths in a farmland as models’ output and meteorological data as models’ input, including mean (Tmean), minimum (Tmin), and maximum (Tmax) air temperatures, evaporation (Eva), sunshine hours (SSH), and solar radiation (SR), which were collected at Isfahan Synoptic Station (Iran) for 13 years (1992–2005). Six different input combination scenarios were selected based on Pearson's correlation coefficients between inputs and outputs and fed into the models. We used 70% of the data to train the models, with the remaining 30% used for model evaluation via multiple visual and quantitative metrics. Our findings showed that Tmean was the most effective input variable for ST forecasting in most of the developed models, while in some cases the combinations of variables, including Tmean and Tmax and Tmean, Tmax, Tmin, Eva, and SSH proved to be the best input combinations. Among the evaluated models, BA-KStar showed greater compatibility, while in most cases, BA-IBK and -LWL provided more accurate results, depending on soil depth. For the 5 cm soil depth, BA-KStar had superior performance (i.e., Nash-Sutcliffe efficiency (NSE) = 0.90, 0.87, and 0.85 for 3, 6, and 9 d ahead forecasting, respectively); for the 50 cm soil depth, DA-KStar outperformed the other models (i.e., NSE = 0.88, 0.89, and 0.89 for 3, 6, and 9 d ahead forecasting, respectively). The results confirmed that all hybrid models had higher prediction capabilities than the LR model.