Improved flood prediction using deep convolutional neural networks in Ellicott City, Maryland

Abstract:

Flooding remains one of the most devastating and costly natural disasters. As flooding events grow in frequency and intensity, it has become increasingly important to improve flood monitoring, prediction, and early warning systems. This is particularly true for places like Ellicott City, MD, a historic Maryland town that has been struck by two “1,000-year” floods since 2016. Recent efforts to improve flood forecasts using deep learning techniques have shown promise, with long short-term memory (LSTM) networks being particularly suited to hydrologic time series prediction. While such techniques are mainly designed to learn temporal relationships, they may not fully capture potentially significant spatial dynamics between input datasets. Here we propose a hybrid approach using a Convolutional LSTM (ConvLSTM) network to predict downstream stage height sequences in Ellicott City, using regional meteorological, hydrological, and land cover inputs. Preliminary results suggest the hybrid network can more efficiently capture the spatiotemporal dynamics of the catchment relative to previous modeling efforts. Furthermore, the improved prediction accuracy and warning times have important implications for local decision making in Ellicott City.