Elephant(s) in the room: mapping neural networks, embeddings and basic models in spatial data science | R bloggers

Slides: https://jakubnowosad.com/agforum2025

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

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

Embedding are compact numerical representations that compress high-dimensional spatial data. They are used for similarity searching, change detection, clustering and classification with minimally labeled data. Google DeepMind’s AlphaEarth Foundations has been producing embeddings of various geospatial data (optical, thermal, radar, elevation, climate) at 64-dimension resolution every year since 2017, without the need for preprocessing. Challenges in using them 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, multimodal alignment, temporal modeling, contrastive learning). Examples include Terramind, AnySat, Prithvi and AlphaEarth Foundations. They produce embeddings, work with minimally labeled data, and can be refined for tasks such as land cover mapping and change detection. TabPFN is a basic tabular data model that can be applied to geospatial predictive maps. Yet foundation models have limited transferability to new regions, and traditional methods remain competitive when labeled data is abundant.

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

@online{nowosad2025,
  author = {Nowosad, Jakub},
  title = {Elephant(s) in the Room: {Graph} Neural Networks, Embeddings,
    and Foundation Models in Spatial Data Science},
  date = {2025-12-15},
  url = {https://jakubnowosad.com/posts/2025-12-15-agforum-talk/},
  langid = {en}
}

Related