Databases like GBIF hold millions of records of where species have been observed over decades. These records are a natural source of labels for pixel classification with the Tessera geospatial foundation model. The difficulty is that each GBIF record often carries some uncertainty about its precise GPS position. This uncertainty is sometimes recorded with the label, and can be 20km+. A classifier that trusts these positions exactly can learn some absurd results, such as elephants swimming miles off the coast of Gabon!
1 Project outline
This project will investigate how to handle GPS uncertainty in GBIF labels when using them as label sources for Tessera tasks:
- The project will first work through GBIF to categorise how much positional uncertainty there is, and how it varies across different parts of the dataset.
- We will then investigate classification methods for training Tessera classification heads that tolerate this uncertainty, so that a label with a position uncertain by many kilometres no longer places a terrestrial mammal in the sea .
2 Approaches
We need a training technique that is aware of geolocation uncertainty. There are a few ways to build one:
- Multiple-instance learning: This draws a circle around each labelled point and assigns every pixel inside the ring to the model as one bag. As it learns, the model pays more attention to some pixels in the bag than to others. This works well with conventional models such as random forests, which are often work fine as heads over the pretrained Tessera embeddings.
- Latent-location models: These treat the true location of each label as something to infer, built around a Gaussian process. This is conceptually close to a hierarchical Bayesian framework above. He et al. take this approach to quantify and reduce the registration uncertainty of spatial vector labels on Earth imagery (KDD 2022), and report good results with it.
The first is the simplest and a good starting point, with the second and as possible extensions. Validation could be done by injecting GPS uncertainty into some known fixed labels and evaluating the effect synthetically.