Using deep transfer learning for Unsupervised image segmentation in remote sensing

Keywords: Convolutional Neural Network, Land Use Land Cover, Deep Learning
Session Type: Virtual Poster Abstract
Day: Friday
Session Start / End Time: 2/25/2022 03:40 PM (Eastern Time (US & Canada)) - 2/25/2022 05:00 PM (Eastern Time (US & Canada))
Room: Virtual 39

Abstract

As multispectral image resolution has increased, generating accurate segmentation of these images can pose a significant challenge. Another hurdle is the presence of accurate labeled data which can require hours of manual segmentation. One solution to this problem is the application of deep learning algorithms which are able to learn non-linear trends in the data without significant preprocessing. Deep learning models can also be used for transfer learning. In this research, we demonstrate transfer learning on how a model trained on one dataset can be used for segmenting a different dataset. We explore the widely known Potsdam and Vaihingen images to achieve our objective. Using a specific deep learning algorithm called UNet, we first train our model on a dataset with class labels. We then use the trained model to extend a custom UNet structure which is able to transfer semantic knowledge from the previous training and also adapt to the unknown images. Preliminary results indicate that there is a potential to achieve higher accuracy by using optimized loss functions suited for unsupervised learning along with pre-trained weights from the trained UNet model.