FLOOD DETECTION MODELING: LEVERAGING THE SEN1FLOOD11 DATASET FOR THE RIO COLIMA RIVER

Climate change has aggravated natural disasters like floods. Floods represent one of the most destructive natural disasters globally, necessitating accurate and timely detection systems. The recent developments in deep learning technology could be leveraged to support flood disaster management as part of climate adaptation. This paper comprehensively evaluates state-of-the-art flood detection models using satellite imagery from the Sen1Flood11 dataset. Five distinct models are evaluated: Vision Transformer (ViT), DeepLabV3 with ResNet-50 backbone, U-Net, VGG16, and Random Forest. The models are evaluated on a dataset comprising 4,383 weakly labeled images covering flood events across 12 countries. The results demonstrate the superior performance of the ViT model, achieving 94.37% accuracy and 88.68% IoU, followed closely by DeepLabV3 ResNet-50 (91% accuracy, 86% IoU) and Random Forest (92% accuracy, 84% IoU). The novel contribution of this work lies in evaluating the generalization capabilities of the flood detection model on unseen data from the Rio Colima River, providing valuable insights into model transferability and practical real-world applicability. The findings contribute to the advancement of automated flood detection systems and highlight crucial considerations for deployment in operational settings toward climate change adaptation.