Unsupervised Change Detection in Satellite Images using Deep Learning

Change detection in satellite image time series is a crucial task for urban growth, environmental change, disaster effects, and land use studies. However, the process of acquiring labelled ground truth data for supervised learning methods is costly, time-consuming, and infeasible for large-scale remote sensing problems. To overcome this shortcoming, this paper presents a novel fully unsupervised deep learning architecture for change detection in a series of multispectral satellite images. The proposed architecture uses a convolutional autoencoder that is trained on a sequence of multi-band satellite images capturing the normal state of a given scene. Each image has 13 spectral bands from the Sentinel-2 satellite image data, allowing for a detailed spatial-spectral characterization. The autoencoder discovers the patterns of unchanged areas without any supervised learning. To enhance the interpretability of the proposed architecture, K-Means clustering is used on the change maps to classify the identified areas into distinct change intensity classes. The proposed architecture is simple, efficient, and free from temporal consistency and supervised learning. The experimental results on multi-temporal satellite image data show that the proposed architecture successfully identifies significant structural and environmental changes with minimal reliance on labelled data.