Temporal analysis of floods using the ConvLSTM network for multisource data classification

Major floods were the most frequent disasters in 2022, accounting for 45% of global occurrences. In Brazil, the number of flood events has been increasing, especially in the state of Rio Grande do Sul (southern Brazil), where 471 municipalities were affected by extreme rainfall between March and May 2024. Remote Sensing (RS) data are essential for managing such disasters, but monitoring flood-prone areas over time is challenging due to the complexity of data processing, topographic variability, and land use changes. Satellite Image Time Series (SITS) enable the analysis of surface water dynamics across different time periods. Deep Learning (DL) models, especially Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs) are effective in extracting temporal and spatial features from geospatial data. This study evaluates the potential of integrating optical, Synthetic Aperture Radar (SAR), and Digital Elevation Model (DEM) data with a dual-branch DL architecture to map flood-prone areas. The model processes dynamic data through a Convolutional Long Short- Term Memory (ConvLSTM) branch and static data through a 2D CNN branch for multi-source data classification. Several feature-level fusion strategies were tested to merge features across branches and over time in Rio Grande do Sul. Among them, the dot-product fusion model achieved the best performance, with a mean Intersection over Union (mIoU) of 0.8207 and F1-score of 0.8862.