Unsupervised deep learning approach to analyze spatio-temporal change in satellite imagery

This paper introduces a deep-learning-based framework to study the change of urban features over time using Satellite imagery. The framework combines an unsupervised convolutional neural network (CNN) with an Incremental K-Means algorithm to capture temporal changes in urban signatures over multiple years. By processing subsets of year-wise satellite imagery datasets, the model incorporates spatial and temporal dimensions, enabling a comprehensive analysis of the study area. The primary focus of this paper is on developing the unsupervised deep learning framework. The framework's ability to identify changes in spatiotemporal clusters of urban features over time lays the foundation for future work. As proof of concept, we present a case study using Five years of satellite imagery and illustrate how to interpret the framework result through visualization. Results show how the satellite tile clusters from satellite imagery change over time, allowing for tracking of gradual urban change and eventual prediction of change into the future.