# Generating potential natural vegetation maps under contemporary climate with Tessera

*2026-08-01 — idea*


Global conservation metrics such as [LIFE](https://anil.recoil.org/projects/life) need a counterfactual baseline to ask questions like "what would the land surface look like today if humans had never converted it to cities, farmland and plantations"?

The usual approach is to use a "[potential natural vegetation](https://en.wikipedia.org/wiki/Potential_natural_vegetation)" (PNV) map. The best available global ones such as [Hengl et al](https://doi.org/10.7717/peerj.5457) and [Jung et al's global habitat map](https://doi.org/10.1038/s41597-020-00599-8) with the [Lumbierres et al crosswalk](https://doi.org/10.1111/cobi.13851), are built from coarse covariates and reflect historic climate conditions.

But a world regenerated or rewilded today would grow back under _today's_ climate, not that of the early twentieth century, and so this project aims to generate PNV maps conditioned on contemporary climate instead.

The idea is to build on our [TESSERA](https://anil.recoil.org/projects/tessera) [foundation model](https://anil.recoil.org/papers/2026-tessera-v2) and the [global habitat mapping](https://anil.recoil.org/ideas/tessera-habitat-maps) work. The world still has large wild regions that are minimally affected by humans, and Tessera gives us a dense 128-dimensional description of every 10m of them. If we mask out the human-affected parts of the planet, the remaining wild pixels form a huge training set relating climate, terrain and soil to what the land surface actually does when left alone. A generative model trained on that relationship could then be run over the masked-out regions to extrapolate what cities and agricultural land would plausibly revert to, expressed as Tessera embeddings that we can decode into habitat classes.

The target output is IUCN-style habitat classes at ~1km, matching what [LIFE](https://anil.recoil.org/projects/life) and area-of-habitat pipelines consume today.
For the climate conditioning, recent gridded climate normals (e.g. [CHELSA](https://chelsa-climate.org) or [ERA5](https://en.wikipedia.org/wiki/ERA5)\-derived) are one starting point, and in future our own [weather downscaling](https://anil.recoil.org/ideas/tessera-weather-downscaling) work could supply localised climate conditioning instead, which pairs Tessera with weather models from the opposite direction.

## Candidate ML approaches

There are several ways to attack the generative step, roughly in order of ambition:

- **Climate-analogue retrieval.** For each human-affected pixel, find wild pixels with the most similar abiotic covariates (contemporary climate normals, elevation, soil) and transfer their Tessera embedding or habitat label. This is essentially the traditional PNV regression, but via  nearest-neighbour search in embedding space.
- **Per-pixel supervised prediction.** Train a model from abiotic covariates to Tessera embeddings (or directly to habitat classes) using only wild pixels, then predict everywhere else in the world. This is the Hengl et al approach, but with Tessera supplying richer training targets than the original vegetation plot data.
- **Masked inpainting over embedding fields.** Treat the human footprint as the mask in a masked-autoencoder setup. First train a model to reconstruct deliberately-masked wild regions from their surrounding spatial context plus climate conditioning, then apply it to the real human-affected mask. This captures spatial ecological structure (riparian corridors, treelines, wetland fringes) that per-pixel models may otherwise miss.
- **Conditional generative models.** A latent diffusion model over Tessera embedding patches, conditioned on contemporary climate and terrain, could sample multiple plausible rewilded surfaces rather than a single answer, giving us an uncertainty distribution over baselines.

Validation is an interesting challenge in itself here, since there's no ground truth for a counterfactual planet. Held-out wild regions give a first reasonable test, as well as long-abandoned agricultural land and older protected areas offer natural experiments. Comparison against the existing PNV maps would locate where contemporary-climate conditioning changes the answer most.

A student wanting to take this project on should be ready to tackle a difficult but rewarding project\!

## See also

The [Global habitat maps from Tessera embeddings](https://anil.recoil.org/ideas/tessera-habitat-maps) project this builds directly upon, the [Mapping LIFE on Earth](https://anil.recoil.org/projects/life) metric that would consume the resulting baseline, and the broader [TESSERA](https://anil.recoil.org/projects/tessera) programme.
Status: Available
Level: MPhil
Year: 2026
Project: TESSERA, a pixelwise geospatial foundation model
Supervisors: Anil Madhavapeddy, Michael Dales, David Coomes, Sadiq Jaffer

## Related

- [Global habitat maps from Tessera embeddings](https://anil.recoil.org/ideas/tessera-habitat-maps) (idea, 2026-08-01)
- [TESSERA v2: Scaling Pixel-wise Earth Foundation Models](https://anil.recoil.org/papers/2026-tessera-v2) (paper, 2026-07-01)
- [Earth embeddings for probabilistic weather downscaling](https://anil.recoil.org/ideas/tessera-weather-downscaling) (idea, 2026-01-01)
- [TESSERA, a pixelwise geospatial foundation model](https://anil.recoil.org/projects/tessera) (project, 2025-01-01)
- [Mapping LIFE on Earth](https://anil.recoil.org/projects/life) (project, 2023-01-01)

---
Canonical: https://anil.recoil.org/ideas/tessera-pnv
Type: idea
Tags: tessera, biodiversity, conservation, ai, spatial, landuse