Semantic Segmentation of Areal Images using Pixel Wise Segmentation

Particularly in regions hit by flooding, precise semantic segmentation of aerial photographs is essential for post-disaster evaluation and urban planning. The high-resolution aerial photography collection FloodNet, created for pixel-by-pixel segmentation of flooded and non-flooded landscapes, is presented in this work. Aerial photos taken from various geographic regions make up the collection, and segmentation masks are used to detect flooded roads, houses, cars, plants, and other vital infrastructure. In order to conduct semantic segmentation on aerial photos and capture the fine-grained information necessary for flood damage assessments, the suggested model makes use of a U-Net architecture. Morphological post-processing is also used to improve object delineation and smooth segmentation borders. The model showed consistent performance with little overfitting, achieving training accuracy of 84% and validation accuracy of 81% over 200 epochs. Effective generalization to unknown data was demonstrated by the validation loss stabilizing at 0.55 and the training loss declining from 1.1 to 0.45. The findings show that the suggested method can reliably identify and categorize areas damaged by flooding from aerial photos, enabling automated flood assessment and speeding up disaster relief activities. There is a lot of promise for integrating this model and dataset combination into real-time catastrophe monitoring, urban planning, and geospatial analytic pipelines.