Change Detection in Satellite Imagery: A Multi-label Approach Using Convolutional Neural Network

Abstract

Change detection in satellite imagery is part of a remote sensing task that aims to take a pair of satellite images of the same region at different times and allocate a binary label for each pixel in the generated change map. A positive label infers that the area depicted by that pixel has changed between the acquisition of the images. Thus, the generated change map has a pixel value as true if a change is detected; otherwise, it is false. The infusion of deep learning in computer vision creates many methods that yield quite impressive results for a single label; however, none of them can deal with multiple labels. Therefore, this paper proposes a multi-label change detection method to differentiate changes based on labels.

In this work, we create a pipeline using a fully convolutional siamese network based on an auto-encoder structure that takes bi-temporal, co-registered satellite images as input and generates a change map as output. The use case of this pipeline is to detect changes on a land cover, e.g., land encroachment, extension of buildings, illegal construction of buildings. As no public multi-label change detection dataset is available, satellite images are hand-annotated to create a new dataset for the task. The obtained results show a 3.9% improvement in the F1-score compared to the current state-of-the-art approach. The state-of-the-art model takes 30 h to train per label, whereas the proposed model takes only 50 h for any number of labels. However, since the Convolutional Neural Network (CNN) based state-of-the-art solution was designed for a single label, the comparison is made on a single label instead of multi-labels.