Identification of Geospatial Patterns using Autoencoders and Clustering

This project demonstrates the effectiveness of unsupervised deep learning for analyzing complex datasets without the need for human labeling. A traditional autoencoder was applied to a dataset of 23 satellite images of US land, leveraging the well-defined grid of the Public Land Survey System to create a uniform dataset. The autoencoder's encoder component compressed the images into a latent space, which was then subjected to K-Means clustering. This analysis successfully identified five distinct geospatial patterns. The study's primary objective was to demonstrate the conceptual transferability of the 'pseudospectrum' concept, previously applied in Raman spectroscopy, to satellite imagery, where the reconstructed cluster centroid is presented as a 'pseudoimage'. This 'pseudoimage' serves as a denoise, idealized archetype of the patterns found within each cluster, proving that this methodology can be successfully applied to discover and formalize hidden patterns in diverse data domains.