Akademik Veri Yönetim Sistemi
Acta Geophysica, 2025 (SCI-Expanded)
This study addresses the challenge of accurately predicting monthly average soil temperature (ST) across Turkey at depths of 5 cm, 10 cm, and 50 cm, aiming to overcome the need for costly empirical measurements. We utilized the ANN-based Elman neural network (ENN) technique, trained with 20 years of data from 77 measurement stations. The model only used geographical parameters—latitude, longitude, altitude, along with periodicity component—bypassing traditional climatic variables. The methodology involved training and testing the ENN model, followed by the use of ArcGIS software and Inverse Distance Weighting spatial interpolation to visualize ST fluctuations and predictions through comprehensive maps. Key findings demonstrate the ENN model’s exceptional precision, evidenced by minimal differences between measured and predicted values, high correlation, and accurate performance across varying depths and seasons. Scatter graphs and regression analyses confirm a strong alignment between predicted and measured data, particularly in training, as shown by the dense clustering around the diagonal line, indicative of precise prediction. In the training phase, the selected model achieved relatively low MAE values (not exceeding 0.9999 °C), RMSE values below 1.2809 °C, and consistently high R2 values not falling below 0.9777. During testing, MAE remained under 1.4410 °C, RMSE did not exceed 1.9435 °C, and R2 values stayed close to 1, with the lowest being 0.9710. Statistical values further affirm the model's overall efficacy despite slight variations at higher temperature ranges, indicating negligible influence on performance. The model accurately identified seasonal ST patterns: highest in Mediterranean and Aegean coastal areas during summer and the lowest in Eastern Anatolia during winter, consistent with real-world observations. This research provides an accurate framework for estimating ST values based solely on geographical data, eliminating the need for direct measurements. The generated maps offer a valuable tool for long-term ST prediction at any unmeasured location in Turkey, significantly advancing our understanding of soil temperature dynamics.