Conference Agenda

Flood susceptibility mapping (FSM) is a crucial component of flood risk assessment; however, traditional statistical and machine learning methods for FSM are limited in their predictive capabilities. FSM approaches typically use static inputs, relying solely on geospatial factors, and fail to consider the spatiotemporal aspects (antecedent conditions) that trigger flood events. This study addresses this gap by developing a hybrid model that combines static geospatial features with dynamic temporal meteorological data, which is often excluded in FSM. The proposed hybrid model consists of two branches: (1) a 2D Convolutional Neural Network (CNN) to extract the features from geospatial inputs (i.e., slope and surficial geology) and (2) a Convolutional Long Short-Term Memory (ConvLSTM2D) network to learn the temporal antecedent conditions from Daymet precipitation, temperature and snow-water equivalent. This model was trained and tested in the Saint John River basin, New Brunswick, Canada — a region that has experienced significant historical flooding. Three hyperparameters were investigated: temporal sequence length (1–4-month timesteps), resampling ratio (0.1-0.7), and positive class weight (1.5 or 2.0). The optimal model was achieved with a 3-month timestep, a 0.2 resampling ratio, and a 1.5 positive class weight, resulting in an F1 score of 0.89. The model performance was highest when using a 3-month timestep, which captured the full snowmelt-to-rain spring cycle, outperforming models that used timesteps of 1, 2, or 4 months. The proposed 2D CNN-ConvLSTM2D architecture is effective in simultaneously learning the static geospatial features and temporal meteorological sequences, highlighting the importance of seasonal antecedent conditions in FSM.

Optical satellite sensors are widely used for rapid flood mapping due to their global coverage and free availability. Thresholding spectral indices, such as the Normalized Difference Water Index (NDWI), can detect water pixels rapidly and with good precision. However, small shifts in threshold values can lead to large differences in flood area and data-driven approaches for threshold selection remain a challenge. At the same time, new foundation segmentation models, such as the Segment Anything Model (SAM), can extract object boundaries from images without task-specific training, though it lacks flood-specific contextual awareness. To address these limitations, we propose a risk-guided segmentation framework that combines risk-weighted optimization of NDWI thresholding, and further refinement of the NDWI mask using SAM. The goal is to improve flood delineation by incorporating information on where a flood is more likely to occur (flood hazard maps) and how flood boundaries appear visually (SAM). We evaluate the method on the 2018 spring flood along the Saint John (Wolastoq) River in New Brunswick, Canada, across five study regions for both Sentinel-2 and Landsat-8 scenes using imagery captured on May 2, 2018 (peak flood for the study regions). We show that a higher risk score corresponds to a higher segmentation accuracy, demonstrating that flood hazard maps can help guide NDWI threshold selection. Moreover, refinement with SAM improves segmentation quality compared to the baseline NDWI masks, demonstrating that the use of risk-guided spectral thresholding with foundation models can improve flood delineation in optical satellite imagery.

Floods remain among the most damaging natural hazards worldwide. In Canada’s capital region—Ottawa and its surrounding areas—flood prediction is crucial, most especially in flood-prone zones, to mitigate recurring events such as the 2017 and 2019 Ottawa floods, which caused extensive damage to homes and infrastructure. This study integrates 18 flood conditioning factors with remote sensing indices and ensemble machine learning to improve flood susceptibility mapping in the Ottawa River watershed. A complementary HEC-RAS 2D hydraulic model simulated flow depth and velocity under a 100-year flood scenario. The ensemble model achieved strong predictive performance (Kappa, F1-score, and AUC > 0.979) and demonstrated high transferability across sub-regions (Kappa > 0.85; F1-score > 0.92; AUC > 0.99). HEC-RAS results indicated spatial variability in flood depth (up to 15 m) and velocity (up to 15 m/s). SHAP analysis identified Elevation, HAND, MNDWI, NDWI, and Aspect as the dominant flood-driving factors. The integrated framework enhances flood susceptibility assessment and supports Natural Resources Canada’s efforts to strengthen flood risk management and resilience in the Ottawa River watershed and similar regions.

Machine learning for flood susceptibility mapping (FSM) has traditionally relied on narrowly scoped events and temporally constrained datasets, limiting the generalizability and long-term utility of predictive models. We present a multi-event, multi-temporal modelling framework that leverages discrete flood occurrences from 2005 to 2023 to train a unified model capable of inference across an extended temporal horizon. Each flood event was treated as a spatio-temporal marker, enabling the model to learn evolving driver–event relationships and underlying temporal trends. Dynamic inputs (e.g., climate data, land use/land cover) are integrated with static geophysical features (e.g., digital terrain model and derivatives) to capture both transient and persistent influences on flood susceptibility. An XGBoost model was trained, tested, and validated using a 70/15/15 split, achieving an overall accuracy of 0.945, with true positive and true negative rates of 0.95 and 0.94, respectively. Precision scores for wet (flood-prone) and dry (non-flood-prone) classes are 0.94 and 0.95. Generated yearly national FSM maps from 2000 to 2023 were evaluated against published flood event datasets. Validation using national flood records, climate variability bulletins, and spatio-temporal analyses of year-to-year raster correlations confirms that years with elevated predicted susceptibility correspond to observed flood events. In addition, a weighted wetness score identified the years with both widespread and extreme flood-prone conditions, highlighting the model’s ability to capture multi-scale temporal dynamics. These results demonstrate that multi-event, multi-temporal modelling enhances the temporal reach and robustness of geospatial flood prediction, providing a foundation for long-term monitoring, trend analysis, and policy-relevant scenario planning.

This study presents a high-resolution geomorphological analysis of river terraces along the urban corridor of the La Silla River (Monterrey Metropolitan Area, Mexico) using UAV-based LiDAR and photogrammetry, with a DJI Matrice 350 RTK equipped with a Zenmuse L2 sensor, generating dense point clouds, DEMs, and orthomosaics. These products allowed for the precise identification of three terrace levels (T1-T3), their geomorphometric attributes, and their lithological composition. The results reveal contrasting degrees of anthropogenic modification: while terrace 1 retains its natural morphology, terraces 2 and 3 show substantial alterations due to residential expansion, public infrastructure, and road construction, which alter the original geomorphological surfaces. Temporal satellite images also show the sensitivity of terrace geomorphology to extreme hydrometeorological phenomena, with cyclones such as Hanna (2020) and Alberto (2024) causing vegetation loss, surface restructuring, and local modification of terraces. Overall, UAV-LiDAR proved to be very effective for mapping terraces in restricted urban environments, providing essential details for monitoring, risk assessment, and sustainable management of urban rivers.