A comparative study of soil temperature models, including machine learning models using two parameters

Abstract

This comparative study examines the efficacy of three established models for predicting soil temperatures at depths of 10cm and 20cm across four Norwegian regions: Innlandet, Østfold, Vestfold, and Trøndelag. To ensure comprehensive regional representation, four monitoring stations were strategically placed within each region. Utilizing data from \acrshort{ac:nibio}, including hourly air temperature at 2m and soil temperatures at 10cm and 20cm depths, the study evaluated seven models cited in existing literature.

These models included Linear Regression, Plauborg’s Linear Regression for daily and hourly values, LSTM, bidirectional LSTM, GRU, and bidirectional GRU. The findings revealed improved performance of bidirectional models over unidirectional ones and comparable results between the hourly extension and Plauborg’s original daily model. Notably, deep learning models exhibited a dual-mode operation to adapt to the transitional Autumn/Spring and stable Summer periods.

It was found that the bidirectional models performed the best and that bidirectional LSTM worked best for 10 cm soil temperature while Bidirectional GRu worked best for 20 cm soil temperature. It was also found that the inclusion of time in regression models improved the models predictive capabilities.

The author of this current study advocates for further research into bidirectional models and suggests broadening the feature set beyond two variables to capture additional predictive variations.