Assessing the spatial variability of soil moisture in the proximity of road networks on a large alluvial fan in the Himalayan Foreland

We use Sentinel-1 and Sentinel-2 images to study drainage congestion due to road networks on a large alluvial fan of the Kosi River. We have estimated the soil moisture from Sentinel-1 images by applying an empirical modified Dubois model (MDM), and a data-driven machine learning model based on the fully connected feed-forward artificial neural network (FC-FF-ANN). We observed that the MDM underestimates the soil moisture (R = 0.43, RMSE = 0.08 m³/m³, and bias = -0.10). The FC-FF-ANN accurately predicts the soil moisture (R = 0.85, RMSE = 0.05 m³/m³, and bias = 0) in our study area. 

We now used the soil moisture obtained from the FC-FF-ANN model to study the spatial pattern of the surface soil moisture in the proximity of road networks that act as drainage barriers. For this, we generated a buffer of 1 km along the road network. Within this, we extract the soil moisture value at the locations where the road network traverses in the vertical, inclined, and horizontal directions. We observed a clear accumulation of soil moisture near the road network that decreases gradually as we move farther from the road. We found that the impact of drainage congestion ranges between 320 to 760 m on either side of the road. This study is a step towards assessing the effect of structural interventions on drainage congestion and flood inundation.