Estimation of soil temperature for agricultural applications in South Africa using machine-learning methods

This study was undertaken to investigate the potential of using machine-learning approaches as alternative and cost-effective tools for estimating soil temperature from readily available meteorological data for agricultural applications in South Africa. Four machine-learning models – multiple linear regression, artificial neural networks, random forest and decision tree – were developed and tested to estimate daily soil temperature at six soil depths (viz. 10, 20, 30, 40, 60 and 80 cm) using meteorological data acquired from seven stations, representing diverse climatic conditions in South Africa. The data were randomly split into two parts: the first 80% of the data set was used for training, while the remaining 20% was utilised to validate the models. The results showed that soil temperature at various depths can be reasonably estimated by different generic machine-learning models, with average Nash–Sutcliffe efficiency values ranging from 0.74 for decision tree to 0.87 for random forest models and root mean square error values of less than 2.79 °C for all models. Among the evaluated models, random forest models showed the highest estimation accuracy across different soil depths and climatic conditions, with average Nash–Sutcliffe efficiency values ranging from 0.87 to 0.95. This study indicated that the performance of climate-specific models was better than that of the aggregated ones. Therefore, it is recommended that machine-learning approaches, particularly RF models, be developed for specific climatic conditions where possible to achieve better soil temperature estimations. The developed models can be applied with caution in other regions with similar climatological and pedological properties. Significance: This study evaluated the performance of four machine-learning models in estimating daily ST at six depths using meteorological data in diverse climatic conditions in South Africa. The results showed that ST at various depths can be reasonably estimated using different machine learning models, although the performance of climate-specific models was better than that of the aggregated ones. Among the evaluated models, RF models had the highest estimation accuracy across different soil depths and climatic conditions.