Deep Learning Approach for Time-Series Prediction of Daily Soil Temperature

Soil temperature (ST) is a crucial climatological parameter with significant applications in solar energy, agriculture, hydrology, geology, agronomy, and environmental studies. Developing accurate models for time-series prediction of ST is particularly important for optimizing agricultural and environmental processes. This study applies a long short-term memory (LSTM) neural network, a deep learning approach, for predicting hourly soil temperature. Hourly ST data from four measurement stations (Adana, Ağrı, İzmir, and Ankara), representing different climatic regions of Turkey, were analyzed over the period 2016–2019. ST values at depths of 20 cm, 50 cm, and 100 cm were estimated using historical ST data. To assess the performance of the LSTM model, statistical metrics such as root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) were used. Results indicate that prediction accuracy improves with increasing soil depth. The lowest MAE (0.0385°C) was recorded at 100 cm depth in Ağrı, while the lowest RMSE (0.0500°C) was observed in Ankara at the same depth, both with the highest R-value of 0.9999. The findings confirm that LSTM is highly effective in capturing the temporal patterns of ST and can be successfully employed in various climatic regions.