Using Geospatial Foundation Models for Species Distribution Modelling
Contact person
Background
Species distributions models (SDMs) estimate the distribution of animal and plant species [1]. They typically are machine learning models that learn the relation between observations of a given species and the environment (land cover, typical weather, terrain, soil conditions, etc). As such, applicability of a particular model is limited to regions where environment data is available. For example, a model trained with specific Norwegian data cannot be used in Sweden.
Geospatial Foundation Models (such as AlphaEarth [2], TESSERA [3], Prithvi [4] or THOR [5]) are neural networks trained on large amounts of satellite imagery as well as auxiliary data such as terrain profile, soil measurements, or climate and weather conditions. They then produce geospatial embeddings that represent a given location. Their geographical coverage is usually at least on the scale of a continent -- far bigger than a single country.
Your Task
Your task will be to investigate whether GFMs can be used to broaden the geographic applicability of species distributions models. Currently, we see two options how this can be done.
We can either (i) extrapolate the input data (by learning the relation between environmental conditions and geospatial embeddings), or (ii) extrapolate the species occurrence (by learning the relation between SDM outputs and geospatial embeddings).
We suggest validating the procedure using geospatially suitable splits in between train and test data (such as checkerboard patterns). If successful, the methodology could also be used to upscale (increase the spatial resolution) of species distribution model outputs.
Prerequisites
Familiarity with the Python ecosystem around machine and deep learning is required. Knowledge about neural networks used in image processing as well as an interest in ecology and remote sensing are advantage.