Where and how can geo-embeddings be used?

From traditional workflows to embedding-driven practice

Aim & Objectives

Aim

  • Understanding how geo-embeddings transform geospatial workflows.

Objectives

  • Explain what geo-embeddings represent
  • Compare traditional and embedding-based workflows
  • Use embeddings for exploration and prediction
  • Critically assess strengths and limitations

Why embeddings matter for geo-practitioners

  • Reduce feature engineering
  • Enable reuse across regions
  • Work with limited labels
  • Faster experimentation

Why now?

  • Rapid growth in EO data
  • Availability of pretrained models
  • Embeddings shared as datasets

Running example

Land cover / land use classification

  • Satellite imagery input
  • Widely used, conceptually difficult
  • Representative of many geo tasks

Traditional LULC workflow

  • Data selection
  • Pre-processing
  • Feature engineering
  • Label collection
  • Model training
  • Post-processing

Where traditional workflows struggle

  • Manual feature design
  • Label dependency
  • Poor transferability
  • Long setup times

Recap: What geo-embeddings are

  • Pretrained image representations
  • Fixed-length semantic vectors
  • Consistent feature space

What geo-embeddings are not

  • Not physical measurements
  • Not directly interpretable
  • Not task-specific

Embedding-based workflow

  • Imagery → embeddings
  • Minimal preprocessing
  • Learning in embedding space
  • Interpretation & validation

Comparing workflows

  • Handcrafted features vs learned representations
  • Task-specific vs reusable
  • Region-bound vs transferable

Application 1: Unsupervised classification

  • Cluster places by similarity
  • No labels required upfront
  • Discover structure first

Why unsupervised matters in practice

  • Handles ambiguity
  • Reveals spatial gradients
  • Supports exploratory analysis

Example outcomes

  • Place typologies
  • Regional patterns
  • Unexpected groupings

Application 2: Predictive models

  • Embeddings as input features
  • Simple downstream models
  • Combine with traditional variables

Why embeddings help prediction

  • Capture complex patterns
  • Reduce model complexity
  • Improve generalisation

What embeddings change

  • Less feature engineering
  • Better transferability
  • Faster iteration

What embeddings don’t change

  • Remove ambiguity
  • Eliminate bias in training data
  • Replace domain knowledge

Common pitfalls

  • Treating embeddings as objective
  • Ignoring training context
  • Over-interpreting clusters

When embeddings add most value

  • Large spatial extent
  • Few labels available
  • Multiple tasks or regions

When traditional features may be better

  • Small local studies
  • Physically interpretable models
  • Policy-driven analysis

Practical considerations

  • Choice of pretrained model
  • Spatial and temporal scale
  • Validation strategy

About Lab II

  • Unsupervised place typology
  • Embeddings as core data
  • Extension to prediction

Key takeaways

  • Embeddings reshape workflows
  • Enable reuse and scalability
  • Require careful interpretation

Looking ahead

  • Foundation models in EO
  • Multi-modal embeddings
  • Future geo-practice