On demand machine learning-driven surface freeze-thaw retrieval across Canadian agricultural regions using Sentinel-1 SAR data

This study explores the prediction of freeze-thaw (FT) states across agricultural fields in four Canadian provinces-Alberta, Manitoba, Saskatchewan, and Québec-using Random Forest classification and regression models. Soil temperature data at a 5 cm depth were gathered from 174 agricultural weather stations from 2016 to 2023. Sentinel-1 SAR VH radar backscatter indicators were processed using Google Earth Engine (GEE). Two modeling approaches were evaluated: a classification model trained on in situ data, where soil states were rigidly classified as either frozen or thawed, and a regression model trained against in situ soil freezing probabilities. Additionally, other site-specific ancillary variables such as latitude, altitude, crop type, and soil type were tested as potential predictors. The regression model using the Exponential Freeze-Thaw Algorithm (EFTA) derived from VH radar backscatter (VH EFTA ) demonstrated strong discrimination between frozen and thawed states, and emerged as the most influential factor, accounting for over 90% of the model’s predictive ability. Models using VH EFTA alone achieved up to 81.4% accuracy for classifying FT state, with only a slight improvement to 82.1% when combined with other predictors. Spatial and temporal validation showed stable accuracy (0.79–0.83) and F1-scores (0.75–0.88) across regions and years. Evaluation of model sensitivity to seasonal and temperature variability revealed that although uncertainties were not fully eliminated during transitional periods for both binary and probability-based FT models, binary-based models consistently showed lower error rates and stronger performance. The final FT model was implemented within an interactive web-based tool that generates on-demand FT maps for user-supplied regions of interest across Canadian agricultural and open-land areas.