Abstract:

 Research on soil temperature prediction mainly focuses on the surface layer, while relatively few studies on deep soil temperature, especially in high−altitude areas where relevant meteorological data are difficult to obtain. Authors employed six machine learning algorithms, which included the traditional Random Forest (RF), Particle Swarm Optimization - based Random Forest (PSO-RF), Ant Colony Optimization - based Random Forest (Ant-RF), as well as three neural network methods [Radial Basis Function (RBF), Backpropagation (BP), and Extreme Learning Machine (ELM)] to predict soil temperature at seven depths (0cm, 10cm, 20cm, 30cm, 40cm, 50cm and 60cm). Meanwhile, comparison of different machine learning algorithms in predicting soil temperatures at different depths were performed. The application of these algorithms in predicting soil temperature at different depths was compared using a dataset of soil temperature and meteorological data from Naqu city (Xizang) from 2017 to 2019. Then, these data were used as input variables for the models, which included temperature, humidity, cumulative solar radiation, precipitation and atmospheric pressure. At the same time, Taylor diagrams were used for model evaluation. The results showed that soil temperatures in the shallow layers changed dramatically, directly influenced by atmospheric temperature and solar radiation, while deep soil temperatures changed more steadily, with certain thermal insulation properties and obvious seasonal fluctuation characteristics. Comparing different models, it was found that in the prediction of soil temperatures at multiple depths, the RBF (radial basis function) neural network model demonstrated higher accuracy in prediction precision, stability and generalization ability, with R² ranging from 0.9016 to 0.9904 and MSE (mean square error) between 0.2501 and 2.7725 °C, achieving the highest accuracy at a depth of 50cm, with R² reaching 0.9904. The model performed best at a 50cm depth, where R² reached 0.9904. Afterwards, the RF model followed, with R² ranging from 0.8861 to 0.9381. Therefore, the RBF model could more accurately capture the complex relationships between soil temperature and various influencing factors, including meteorological conditions and soil depth. This study provides a more reliable and accurate tool for predicting soil temperatures at different depths, thereby offering important scientific basis for fields such as agricultural management, environmental protection, and climate change research.

Key words: Soil temperature, Machine learning, Neural networks, Forecast, Random forest, Xizang Naqu