Embedding Locations with Satellite Image Encoders

By Levien van Krieken

Abstract
Traditionally, locations on Earth are mainly compared spatially. Recent advancements in deep learning encoding models now allow comparison between embedded locations instead. These embeddings are high-dimensional vectors (256-2048 dimensions) that encode semantic features of locations. This thesis considers peninsular Spain and compares different strategies for encoding the locations within. Most strategies rely on image encoder models which extract location-specific image features from multi-spectral Sentinel-2 satellite imagery. These deep learning vision models are pre-trained with self-supervised learning algorithms allowing for feature rich embeddings of the locations without the need for any training. The embeddings were analysed qualitatively by comparing the semantic embedding similarity to the spatial similarity. Quantitatively, the embeddings were evaluated on various downstream prediction tasks correlated to human diseases. The research compares encoding models with diverse feature complexities and examines their correlations with environmental disease predictors. Additionally, it provides recommendations on future applications of embedded locations for disease mapping.