Utilizing Improved Model Validation Techniques to Reevaluate the Sensitivity of Monthly Temperature Predictions to Multiple Sources of Soil Moisture Data

Soil moisture can improve temperature forecasts at subseasonal-to-seasonal (S2S) timescales, particularly during the warm season. To account for the growing availability of soil moisture data, this research quantifies the sensitivity of statistical forecast models to three different soil moisture datasets across the contiguous United States (CONUS). This study builds upon previous work which integrated soil moisture data from direct in situ measurements, land surface models, and proxy estimates into forecasts made using quantile regression models to examine how the source of data impacts model performance. A limitation of this method was that data sensitivity was only tested using in-sample model validation. Therefore, this study uses a more rigorous selection of in situ soil moisture sites to conduct out-of-sample model validation and reevaluate the sensitivity of predictions to the soil moisture data source. Results show that differences in model performance remain small and that in most cases the statistical models are not sensitive to the type of soil moisture used. The shallowest soil moisture measurements displayed the strongest relationship with temperature and the soil moisture proxy performs similar to the shallow soil moisture layers. The results of this study also demonstrate that soil moisture is shown to be a skillful predictor of warm season monthly temperatures out to a two-month lead time, or within one season. There was a significant decrease in forecast skill as the lead time increases but results also showed that the percentage of locations with higher skill using soil moisture increases at longer lead times. This highlights forecasts of opportunity, where soil moisture’s value over statistical forecast baselines occurs at one- and two-month lead times.