Soil moisture forecast based on gridded historical and forecast datasets

Continuous monitoring of soil moisture (SM) has become a prevalent approach in precision irrigation control. Fluctuations in SM within the root zone, whether caused by overly wet or dry conditions, can potentially diminish plant transpiration, leading to decreased productivity. Hence, ensuring a timely and appropriate supply of water is essential for effective irrigation management. Though various machine and deep learning models, along with in-situ climate data, have been examined for monitoring SM, the incorporation of gridded historical and forecast climate data into this aspect has not been explored. In this research, we assess forecasting SM by Random Forest (RF) model for the next 7 days using two approaches: A) relying on forecasted data for each day, and B) relying solely on historical data. To this end, the gridded climate data (air temperature, relative humidity, wind speed, precipitation, and reference evapotranspiration-ET0), the soil features (lagged in-situ SM and gridded soil temperature), and vegetation features (Normalized Difference Vegetation Index-NDVI) for different land covers in Oulu, Finland. The findings suggest that using gridded data could be a promising option in places where there is limited data for the SM forecasting. The lagged SM was the most explaining variable, followed by soil temperature, NDVI, and ET0. Furthermore, both scenarios exhibited similar trends, showing a decline in forecasting accuracy as the lead time approached 7 days, and thus scenario B can provide more efficient SM forecasts.