The sheer volume of Earth Observation (EO) data today—a planetary deluge of multi-spectral, radar, and LiDAR feeds is overwhelming traditional remote sensing techniques. For decades, analysts relied on meticulous, band-by-band spectral science, but this approach is simply unsustainable at scale of “new space”.
Enter the hot term of the moment, geospatial embeddings, a core innovation of modern Artificial Intelligence. These high-dimensional vectors are now becoming a foundation of the field, promising to unlock insights at planetary scale.
However, while we embrace this powerful new capability, we must also critically examine the sacrifice of transparency and trust in favour of efficiency.
What Are Geospatial Embeddings, and What Do We Lose?
At its most fundamental, an embedding is a data compression and transformation technique. It converts highly complex, multi-temporal EO data— both the spectral and contextual history of a single spot on Earth—into a compact, continuous, numerical vector.
This is achieved using massive Foundation Models trained on trillions of pixels globally. These models it’s claimed extract the “semantic” meaning of the data, not just what a location looks like right now, but what it is (e.g., ‘a coastal mangrove forest’) and how it behaves (e.g., ‘undergoing gradual drought stress’).
The claims are indeed impressive: similarity of landscapes in the real world equals proximity in the vector space.
However, the sacrifice is the direct physical interpretability that traditional remote sensing was founded on. A simple Near-Infrared value tells you about leaf density, while an embedding dimension tells you… nothing on its own.
We supposedly gain “semantic” meaning, but we trade it for algorithmic opacity.
I appreciate this might me difficult to follow, perhaps a little nostalgia to my formative years in Remote Sensing might help..
Back to the future.. Image processing 1980’s style !
When I was a boy, or at least a Masters Student of Applied Remote Sensing, when it came to segmenting an image or classification to use the term of the day, you had the choice of using special signatures at various wavelengths to train a classification algorithm – supervised classification or using algorithms to identify similar groups of pixels automatically – a technique know as unsupervised classification.
So the concept of grouping similar features isn’t new. For decades, analysts relied on unsupervised classification using algorithms like K-Means. However I would argue although less transparent than supervised classification, unsupervised classification was still superior to embeddings in terms of explainability.
A comparison reveals how fundamentally the connection between the analyst and the data has changed.
| Feature | Geospatial Embeddings (AI Paradigm) | Traditional Unsupervised Classification (Spectral Paradigm) |
|---|---|---|
| Source | Learned Feature Space. Features are abstract, emergent properties generated by a complex neural network. | Spectral Space. Features are direct, measurable spectral values or simple ratios (e.g., NDVI). |
| Interpretability | Semantic but Opaque. The meaning is distributed across hundreds of dimensions, making it a “black box” feature. | Physical and Direct. Clusters can be traced back to raw spectral values and physically understood. |
| Consistency | Globally Consistent. Highly efficient but requires trust in a model trained by others on proprietary datasets. | Locally Specific. Requires intense manual effort to label, but the process and output are fully transparent and auditable. |
| Cost of Error | Silent Failures. If the global model fails in a niche local context (e.g., high-altitude agriculture), the error is baked into the compressed vector and difficult to debug. | Traceable Failures. Errors are usually traceable to sensor issues, atmospheric conditions, or poor cluster separation, allowing for immediate intervention. |
The move to embeddings offers a necessary efficiency gain, but we should acknowledge that we are embracing a system that outsources the core act of feature engineering to a black box. While unsupervised classification was cumbersome, its output (a handful of spectral clusters) was fully auditable and understandable by a human analyst.
Embeddings are powerful precisely because they are cryptic.
Spatial Autocorrelation Ignoring the law of Geography ?
Perhaps the greatest internal challenge facing the adoption of geospatial embeddings lies in the very nature of geographic data itself: spatial autocorrelation.
Spatial Autocorrelation is the formal expression of Waldo Tobler’s First Law of Geography: “Everything is related to everything else, but near things are more related than distant things.”
In the context of Earth Observation, this means that a pixel representing a healthy forest has an attribute (its spectral signature and, consequently, its embedding vector) that is highly similar to the pixel immediately next to it. Features on Earth—like a large field, a mountain, or a lake—don’t stop abruptly at a single pixel; they extend continuously across space.
This of course sounds logical, but for AI models this is both a blessing and a curse..
The Blessing (Efficiency): Deep learning models, particularly Convolutional Neural Networks (CNNs) used to generate embeddings, thrive on spatial autocorrelation. Their convolutional filters are designed to look at a neighbourhood of pixels to determine the context of a central pixel, effectively enforcing Tobler’s law by ensuring that the resulting embedding vector captures local spatial dependencies. This is what makes the embeddings so rich in context.
The Curse (Statistical Integrity) : Spatial Autocorrelation poses a significant challenge to statistical integrity in geospatial data. In classical statistics and machine learning, we assume data points are independent and identically distributed. However, this assumption is often violated in geospatial data particularly when we combine spectral EO data with other contextual data such as elevation or temparture data. This leads to two critical issues in embedding-based workflows:
- Inflated Model Performance: When models are trained or validated using samples too close together due to spatial autocorrelation, the high similarity results in overly optimistic performance metrics. The model doesn’t learn a generalisable rule but rather memorises local spatial trends, leading to the false belief that an embedding is highly predictive when applied to distant, dissimilar regions (spatial heterogeneity)
- Sampling Bias: Analysts must carefully de-correlate their samples when creating ground truth labels for downstream models (e.g., classifying land cover based on pre-computed embeddings). If samples are taken too close together, they contain redundant information, skewing the training process and failing to capture the true global variability of the phenomenon.
Show your workings..
Remember exams when you were asked to explain out you got to you answer…
An embedding is a point in a vast, high-dimensional space:
If an AI model, using this vector, flags a specific region as “High Risk for Illegal Deforestation,” policy makers and conservationists need to know why. What specific real-world feature does the value of -0.81 in the 32nd dimension correspond to?
The blunt answer is: it doesn’t correspond to any single thing. The intelligence is collective.
This has all the makings for a crisis of trust in politically sensitive applications of geospatial information in particular at a global scale..
- The Scientific Audit: Unlike traditional spectral indices, embeddings cannot be easily validated using ground truth data due to their nonlinear and opaque input-output relationship. This makes it challenging to verify the scientific rigour of insights based on untraceable fundamental feature representations.
- Bias and Fairness: If the foundational model was primarily trained on data from the Global North, its embeddings might poorly or unfairly represent landscapes, agricultural practices, or built environments in the Global South. This algorithmic bias is difficult to detect and correct within the black box of the vector space.
- Regulatory Compliance: As governments increasingly rely on EO-derived intelligence for policy and enforcement, a critical question arises: can an AI prediction based on an uninterpretable embedding vector withstand legal scrutiny? The requirement for a human-understandable chain of evidence remains a significant roadblock.
Conclusion
The path to reliable, globally robust geospatial AI lies not just in bigger models, but in models that respect, measure, and account for the fundamental geographic laws that govern our planet.