Segmentação semântica de inundações em imagens de satélite por meio de redes neurais totalmente convolucionais

Abstract

Natural disasters pose a global challenge, causing significant damage across various countries. Floods are extreme events leading to loss of lives, infrastructure damage, and substantial economic losses. Remote sensing through radar offers valuable insight into identifying flooded areas due to its sensitivity to water presence. Artificial neural networks have successfully been used to detect floods in radar images, thanks to their ability to learn complex patterns within these images. This study proposes the development of a neural network utilizing U-Net and U-Net++ architectures to semantically segment flooded areas in radar images. To optimize segmentation accuracy, various loss functions such as Binary Cross-Entropy, Dice, and Tversky (α = β = 1) were explored. The experiments indicated that U-Net++ with Binary Cross-Entropy stood out when analyzing the predictive results of the models.