Authors: Sadra Seyvani – University of Alabama
Title: Deep-Learning-Based Benchmark Flood Inundation Mapping Using RGB Aerial Imagery
Presentation Type:
Abstract: High-resolution flood inundation maps (FIMs) are essential for studies in hydrological applications. One of the most conventional methods to access the benchmark FIMs is the use of high-resolution aerial imaging. Using the drone-captured aerial imagery, we can avoid common limitations of satellite imagery, including cloud cover, low spatial resolution, and a lack of temporal consistency. Besides the advantage of high accuracy of the mentioned dataset, the spectral content of the dataset is limited to the red, green, and blue (RGB) bands, which prompted insufficient use of the conventional remote sensing approaches, necessitating a more advanced methodology. To address these issues, we introduce a novel neural-network architecture called Sub-Matrix Convolutional Neural Network (SMCNN), which classifies each pixel based solely on its centered spatial neighborhood (e.g., 256×256 patches), avoiding the asymmetric predictions of conventional encoder-decoder models. Unlike common neural-network architectures like U-Net, the SMCNN ensures consistent spatial treatment across both central and edge pixels, enhancing classification accuracy near flood boundaries. Preliminary results, even after a single epoch of training, demonstrate the strong potential of the SMCNN, achieving 87.6% accuracy and 90.3% recall, outperforming the U-Net in recall despite limited training. These results highlight the capability of the proposed method to capture fine-scale spatial patterns of flood extents using RGB data, offering a scalable benchmark solution for future flood modeling research.