Elephant(s) in the room: Graph neural networks, embeddings, and foundation models in spatial data science

Slides: https://jakubnowosad.com/agforum2025

This presentation covered three interconnected deep learning concepts appearing in spatial data science work.

Graph Neural Networks (GNNs) are a deep learning architecture that represents spatial data as graphs: nodes are spatial units (pixels, regions, locations) and edges are relationships (proximity, similarity, connectivity). Nodes aggregate information from neighbors through message passing, similar to spatial lag models. Common types include Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), GraphSAGE, and Graph Isomorphism Networks (GINs).

Embeddings are compact numerical representations that compress high-dimensional spatial data. They’re used for similarity search, change detection, clustering, and classification with minimal labeled data. Google DeepMind’s AlphaEarth Foundations produces 64-dimensional, 10-meter resolution embeddings from diverse geospatial data (optical, thermal, radar, elevation, climate) for every year since 2017, requiring no preprocessing. Challenges in their use include interpretability and selecting appropriate embeddings for specific tasks.

Foundation models are large pre-trained models that learn general representations from massive unlabeled Earth observation data through self-supervised learning (masked image modeling, multi-modal alignment, temporal modeling, contrastive learning). Examples include Terramind, AnySat, Prithvi, and AlphaEarth Foundations. They produce embeddings, work with minimal labeled data, and can be fine-tuned for tasks like land cover mapping and change detection. TabPFN is a foundation model for tabular data that can be applied to geospatial predictive mapping. Still, foundation models have limited transferability to new geographies, and traditional methods remain competitive when labeled data is abundant.

During the talk, I explained the basics of these concepts and, for each one, showed practical applications in R using reproducible examples. They included landform classification with GNNs, change detection using AlphaEarth embeddings, and species richness mapping with TabPFN.