Learning from Geographically Biased Spatial Data

43 concepts40 papers1 proposal

Imagine trying to map the distribution of a rare plant species using crowdsourced citizen science observations. If you plot the raw data points on a map, you will quickly notice a striking pattern: the observations line up perfectly along hiking trails, highways, and urban parks.

A naive machine learning model trained on this data would conclude that the plant thrives on asphalt and avoids remote wilderness. This is the core challenge of spatial sampling bias. Because human data collection is opportunistic—driven by convenience, population density, and infrastructure—our datasets rarely reflect the true underlying distribution of geographic phenomena.

Standard machine learning algorithms assume that training data is independent and identically distributed (i.i.d.). Spatial data routinely violates this assumption. As noted by Carla P. Gomes et al., large-scale environmental monitoring datasets suffer from highly irregular sampling densities and skewed distributions. This issue spans multiple domains: Peter Dayan et al. observe that even biological brains must find ways to build stable spatial maps despite an animal's non-uniform exploration of its environment, while Tim Alpherts demonstrates that global street-view datasets heavily over-represent frequently traveled urban routes rather than providing a uniform mirror of our cities.

When models treat the absence of an observation as a true negative, they end up modeling human geography rather than ecological or physical reality. To build models that generalize to under-sampled regions, researchers have developed mechanisms that decouple observer effort from true underlying patterns.

Disentangling Effort from Occurrence

The fundamental dilemma of crowdsourced spatial data is that it is often "presence-only." We know where an entity was spotted, but we rarely know whether a blank space on the map means the entity is truly absent or if simply no one looked there.

One elegant solution to this problem is the target group background approach. The intuition is simple: instead of comparing a specific species' presence against random points on the globe, the model compares it against the observations of all related species recorded by the same group of observers.

As implemented by César Leblanc et al., this approach allows the model to quantify observer effort. If a particular geographic region contains thousands of bird sightings but zero sightings of a specific hawk, the model can safely infer a true absence. Conversely, if a region contains zero sightings of any bird, the model recognizes a lack of effort rather than a biological void, preventing it from penalizing predictions in data-poor zones.

Naive Approach:
[Data Point] ──> Treated as Presence
[Unsampled Wilderness] ──> Treated as Absence (Error!)

Target Group Background Approach:
[Unsampled Wilderness] ──> No observations at all ──> High Uncertainty / Ignore
[Highly Traveled Trail] ──> 1,000 total observations, 0 Target Species ──> True Absence

When verified absence data is available but extremely scarce, semi-supervised frameworks can bridge the gap. Tai Peng et al. demonstrated a pseudo-labeling framework that combines a small set of systematically surveyed presence-absence records with a massive, biased presence-only dataset. By using the clean, unbiased data to anchor the initial predictions, the model can safely generate pseudo-labels for the unmonitored regions, effectively cleaning the crowdsourced data during training.

Enforcing Spatial and Ecological Continuity

If human observers do not visit remote regions, how can a model learn what exists there? The key is to leverage the continuous nature of physical space. Geographically close locations tend to share similar environmental characteristics—a principle known as spatial autocorrelation.

Johan Björck et al. exploit this continuity using spatial contrastive learning for remote sensing imagery. Instead of relying on human-provided labels, the authors train a neural network to recognize that geographically adjacent patches of land are "neighbors" that should share similar representations in an embedding space. Because the model learns the smooth, underlying transitions of the landscape (such as changes in soil or vegetation) entirely through self-supervision, it remains robust even when the downstream labeled data is highly clustered around accessible roads.

Spatial Contrastive Learning Intuition:
[Sampled Patch A] <──(Close Geography)──> [Unsampled Patch B]
       │                                         │
       ▼                                         ▼
Forced to have similar embeddings ──> Generalizes features across data gaps

Taking this concept a step further, Elijah Cole et al. introduced Spatial Implicit Neural Representations (SINRs). Instead of dividing the globe into discrete grid cells or treating locations as independent samples, the authors train a deep neural network to map raw coordinate inputs $(x, y)$ directly to environmental and species presence embeddings.

Because deep neural networks possess an inherent inductive bias toward learning smooth, continuous functions, the SINR architecture naturally interpolates across data-sparse gaps. The model cannot easily create sharp, artificial "cliffs" in probability space just because a hiking trail ends; instead, it smoothly extends its predictions into the unvisited wilderness based on the broader geographic context.

Leveraging Multi-Entity Anchoring

When modeling a rare or under-sampled entity, individual data models easily overfit to the specific places that entity was observed. However, spatial entities do not exist in isolation; they share environmental niches and constraints with well-sampled entities.

Philipp Brun demonstrated that modeling thousands of species jointly within a single deep learning architecture yields significantly more accurate community maps than training individual models for each species. In a joint architecture, the dense, high-quality spatial data of common species acts as a regularizer. The shared hidden layers of the network learn a robust mapping of the underlying environment, which anchors the predictions for rare species and prevents the model from overfitting to the biased paths of opportunistic observers.

Similarly, Christian Lange et al. developed an active learning framework that treats the unmapped range of a data-scarce species as a weighted combination of the known ranges of thousands of well-documented species. By framing the problem as a combination of established spatial primitives, the authors show that models can achieve high precision with a fraction of the field observations, bypassing the need for uniform global sampling.

Key Foundational Work

A familiarity-based learning procedure for the establishment of place fields in area CA3 of the rat hippocampus (2000) — contributed 1 concept elements to this direction.
eBird: A Human/Computer Learning Network for Biodiversity Conservation and Research (2012) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_geographic_sampling_bias_spatial_samplin_61fc6897,
  title = {geographic sampling bias (spatial sampling bias)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geographic-sampling-bias-spatial-sampling-bias/rdr_61fc68975b1546a6acca7cd2121679e0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_geographic_sampling_bias_spatial_samplin_61fc6897_v1,
  title = {geographic sampling bias (spatial sampling bias)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geographic-sampling-bias-spatial-sampling-bias/rdr_61fc68975b1546a6acca7cd2121679e0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake