Predicting Missing Soil Temperature Records at Two Forest Sites in England, with Machine Learning

Soil temperature records at forest monitoring sites often contain gaps. Sensors and data loggers fail, and the missing hours must be filled. We compared three methods for this task: climatology, linear regression and random forest. The data were hourly soil temperatures at 0.3 m depth from two sites of the International Co-operative Programme on Assessment and Monitoring of Air Pollution Effects on Forests (ICP Forests) in southern England. The records covered 2022 and 2023, with 12,848 and 17,136 usable hours. Continuous artificial gaps of one day, one week and 30 days (about one month) were removed one at a time. The methods were also tested on the final 20% of each record. Random forest gave the lowest mean absolute error (MAE) for gaps of one day at one site (0.42 °C). For gaps of one week, the two regression methods performed alike. For gaps of one month, linear regression was more accurate at both sites (MAE 0.57 and 0.78 °C), and it was also best on the final part of the records. Soil temperature at one site was predicted from the record of the other site, 244 km away, with an MAE of 0.58 °C. The model was calibrated on the hours in which both sites were recording. The reverse direction gave 0.56 °C. This linear model filled a real gap of almost five months caused by a logger fault. Its approximate prediction intervals came from an artificial outage in the same season of 2023. Moving averages of air temperature reduced the errors of both models clearly. Simple and free methods can thus fill gaps at the two studied forest sites.