Using Temporal Graph Neural Networks to Leverage High Fidelity Observation Data for Improving Generalizability of Hydrological Models
The incredible performance of Deep Learning (DL) has helped make progress in widespread applications, including computer vision, facial recognition, and natural language processing. The success of DL in the geosciences have been limited due to challenges related to data sparsity, complex multi-scale dynamics, and spatiotemporal dependencies. In this paper, we explore a novel approach to combine DL and Soil & Water Assessment Tool (SWAT) model analysis, to improve hydrological forecasts and help guide water management and conservation activities. The proposed framework uses Temporal Graph Convolution Neural networks (TGCNs) to resolve complex hydrological response in a domain consisting of 3000 watersheds. Hydrological Response Units (HRUs) were clustered in a representative feature space using dynamic time warping (DTW). The clusters and DTW distance was used to construct graphs that capture the relationship and degree of connectivity between catchments with similar hydrological characteristics. This framework is used to forecast soil moisture and evapotranspiration in a case study in NorthEast of the US. SWAT was used to simulate field-scale time series data spanning 40 years. The data includes forecasts of key variables such as soil moisture, water balance, evapotranspiration, plant growth, and yield. The simulated data was split into training, validation, and testing data sets for TCGN. Observation data is available for approximately 15% of the watersheds and was used to update predictions at the remaining locations. The TGCN framework was then deployed for the 3000 watershed and the ability to predict soil moisture and evapotranspiration was evaluated. Results were compared against both observation data and the SWAT forecasts. In particular, the ability to generalize to catchments with different hydrological response were explored and characterized.