Harnessing CNN for flood mapping: Insights from Landsat-8 imagery

Flooding is a global natural disaster that causes infrastructures and environmental damages as a result of rapid urbanization, increasing population density, and climate change. In Morocco these events are still causing significant environmental and economic damage, especially in mountainous regions and densely populated areas. Our study focuses on the Zat sub-basin among the Tensift Basin, and on flood mapping using a practical method that utilizes freely available (Landsat-8 satellite imagery) processed through (Google Earth Engine) (GEE). We applied a U-Net Convolutional Neural Network (CNN) model trained with (pre- and post-flood images), and complemented by elevation data from SRTM dataset, to identify flooded inundated areas. We used the automatic thresholding method OTSU for the training data and to distinguish water surfaces from dry land. To assess the model's ability and recognize flood patterns, we used standard metrics such as (overall accuracy), (Intersection over Union) (IoU), and (F1-score). The trained network reached a precision close to 96%, an F1-score of 0.84, and an IoU of 0.73, demonstrating reliable detection of flooded areas in complex topographic settings such as the Tensift Basin.