Deep learning interpretability for transfer learning in drought prediction

Droughts are devastating natural disasters that are becoming more extreme and spreading to new locations as climate change is progressing. Catchment streamflow is a leading indicator of drought conditions. Our work focuses on developing machine learning methods for streamflow prediction. Deep learning methods require large amounts of labeled data for training. Adequate sizes of training data sets are easily collected in well-gauged catchments, but are not available in poorly gauged catchments. Therefore, an aspect of this project is to study the transferability of deep learning models between well-gauged and poorly-gauged catchments. We use both static catchment characteristics from GSIM and dynamic GLDAS data to train an LSTM model and evaluate the ability of the model to generalize to new catchment regions via transfer learning. We also explore model interpretability using standard attribution techniques to identify the connection between input features and streamflow prediction. We are in the process of developing baseline models, both random forest and a simple LSTM, for initial results and will then build a more complex LSTM and incorporate the interpretability assessment of the model. We also plan to explore transfer learning techniques typically used in other deep learning applications to increase the robustness of our model to new regions and changing behaviors in the climate.