Add to Chrome
Add to Firefox
Add to EdgeExploiting ConvNet Diversity for Flooding Identification
Jun 05, 2018
Flooding is the world's most costly type of natural disaster in terms of both economic losses and human causalities. A first and essential procedure towards flood monitoring is based on identifying the area most vulnerable to flooding, which gives authorities relevant regions to focus. In this work, we propose several methods to perform flooding identification in high-resolution remote sensing images using deep learning. Specifically, some proposed techniques are based upon unique networks, such as dilated and deconvolutional ones, while other was conceived to exploit diversity of distinct networks in order to extract the maximum performance of each classifier. Evaluation of the proposed algorithms were conducted in a high-resolution remote sensing dataset. Results show that the proposed algorithms outperformed several state-of-the-art baselines, providing improvements ranging from 1 to 4% in terms of the Jaccard Index.
