Integration of Temporal Meteorological and Geospatial Data for Flood Susceptibility Modelling in Canada using a Hybrid CNN-ConvLSTM Model

Flooding is the costliest disaster in Canada, yet traditional flood susceptibility modelling is computationally expensive for large-scale applications and often relies on static geospatial features while excluding temporal antecedent conditions. This study uses the Canadian Flood Archive maintained by Natural Resources Canada (NRCan) to develop a large-scale flood susceptibility model for Canada's major drainage areas. The model integrates static terrain derivatives, dynamic climate variables, and semi-static geospatial variables using a hybrid Convolutional Neural Network-Convolutional Long Short-Term Memory (CNN-ConvLSTM) framework. The Canadian Medium Resolution Digital Elevation Model (MRDEM) was used to derive geospatial features, including height above nearest drainage (HAND), Euclidean distance to rivers (EUC), slope, aspect, topographic position index (TPI), and terrain ruggedness index (TRI). Semi-static geospatial variables include land cover (available every 5 years) and the annual normalized difference vegetation index (NDVI), which were temporally matched to each historical flood event. The static and semi-static features were coupled with Daymet meteorological data (precipitation, temperature extremes, snow water equivalent) spanning 1–3-month antecedent windows. The performance of the 2D hybrid CNN-ConvLSTM model will be compared with an Extreme Gradient Boosting (XGBoost) baseline. While XGBoost has performed well in prior research, the hybrid CNN-ConvLSTM is hypothesized to offer superior interpretability of flooding mechanisms. By leveraging the temporal sequence of meteorological drivers, the model captures complex spatiotemporal dependencies that traditional machine learning methods cannot. A preliminary sensitivity analysis of temporal sequence lengths (1-3 months) and resampling ratios (0.1-0.7) showed that the CNN-ConvLSTM architecture achieved the highest predictive accuracy (F1 = 0.89) with a 3-month sequence length and a resampling ratio of 0.5. These initial findings suggest that capturing the full spring snowmelt-to-rainfall cycle is critical for flood susceptibility mapping in Canadian watersheds.