Short term hydrological forecasts using surface soil moisture observations from Sentinel-1 images

Abstract:

Imaging satellite radars have shown to be effective at retrieving soil moisture. However, the approach is hampered by factors such as vegetation that masks partially or totally the contribution of the underlying soil moisture to the total backscattered signal. Also, radar images can only retrieve soil moisture over the first few centimeters of soil, while runoff is affected by soil moisture over larger depths. Thus, the potential of satellite derived soil moisture from radar data for flow forecasts has yet to be demonstrated.

27 Sentinel-1A synthetic aperture radar (SAR) images spaced 12 days apart were acquired during the 2018 and 2019 summer seasons over Au Saumon and Magog watersheds in Southern Quebec, Canada, with areas of 1022 and 1764 km2. Watersheds are mainly forested, with agricultural fields representing respectively 9 and 17% of total surface area. 34 and 10 soil moisture probes respectively at 5 and 20 cm depths were deployed on the watersheds.

A linear model to retrieve surface soil moisture from SAR data was developed for low (< 20 cm) and high (> 20 cm) vegetation using in situ soil moisture measurements and radar backscattered signal. Extrapolation from surface to deeper soil moisture was performed using a conceptual model calibrated using in situ soil moisture data. A distributed hydrological model was calibrated on each watershed and hydrological simulations were carried out by updating model state with an optimal weighting of SAR and model derived soil moisture based on historical SAR soil moisture observations and model runs.

The SAR soil moisture model was able to retrieve surface soil moisture with RMSE of 0.030 and 0.033 m3/m3 and R2 of 0.52 and 0.22 for low and high vegetation, respectively. Extrapolation from surface to deeper soil moisture produced RMSE of 0.029 and 0.034 depending on soil type. The soil moisture updating procedure in the hydrological model reduced RMSE by 19% between observed and simulated flows, although the updated model completely missed some runoff events, particularly for lower peak values. Unsurprisingly, the watershed with the largest fraction of agricultural/pasture areas performed better. Finally, the performance of the updated model using only in situ soil moisture reduced RMSE by 22%, therefore performing better than updating the model with Sentinel derived soil moisture.