A global map learns which land may suit solar
Global Solar Siting from Geospatial Embeddings · arXiv:2611.08831
In plain words
Finding land for solar is difficult because geographic scoring rules chosen by experts may not work equally well in different regions. This paper introduces a method and a global map that score land for solar suitability.
The method learns from land data recorded before existing solar facilities were built. In tests across thirteen countries, it ranked future solar sites above random land, though the strength of that ranking varied by country.
The research
Choosing where solar could fit depends on geographic signals that expert-built scoring rules may not carry across borders.
A global suitability map could change which places merit closer attention by showing whether land patterns learned from existing installations can rank future sites across countries.
The method combines a random forest and a neural network to turn global land embeddings into pixel-level solar suitability scores.
The method learns from land embeddings recorded before existing plants were installed, then checks scores against later installations and random land across countries. Future sites generally ranked higher, and the ensemble beat reconstructed expert-weighted scoring in every tested country. This supports a transferable suitability signal within those evaluations.
The method learns from existing solar facilities, which it assumes are well sited. It scores individual pixels, so it does not assess whether enough connected suitable land is available for a plant or account for future climate changes.
What the evidence shows
- A global signal from land
The global solar suitability method is a mapping method that learns from geospatial embeddings, which summarize properties of land, and releases a dataset scoring land at 10 m resolution. This addresses a problem with common siting approaches: they combine hand-selected geographic features with expert-chosen weights, and many are limited to a country or region.
The method asks whether patterns learned from existing solar sites can help rank land across countries.
- A stronger pixel ranking
On held-out test pixels, the ensemble scored 0.932 ROC-AUC, compared with 0.702 for a prototype that ranks pixels by similarity to the average solar embedding. ROC-AUC measures how often a randomly selected solar pixel ranks above a randomly selected non-solar pixel.
The contrast shows that the ensemble extracts a stronger ranking signal than this simple similarity method on the pixel test.
- The signal carries forward
The forward check compared future solar installations with random land in each country. Across thirteen countries, ROC-AUC ranged from 0.83 to 0.95, with higher values indicating that future sites more often ranked above random land.
The range varies by country. This result supports the presence of a useful signal beyond the sites used for training, within the tested countries and period.
- Against reconstructed expert scores
On the same site-versus-random task across thirteen countries, the ensemble exceeded the strongest of sixteen reconstructed MCDA configurations in every country. The ROC-AUC advantage ranged from 0.058 to 0.299 points.
This comparison favors the learned method over those expert-weighted configurations, while leaving other siting methods outside its scope.
- What the global map captures
The global ranking connects suitability scores to existing deployment: the most-suitable fifth of land contains 67% of installed capacity and 84% of estimated annual generation. These are observed capacity and modeled generation associated with land ranked by the method, not guarantees about future plants.
Since the model learns from existing facilities and scores pixels, its map is a prioritization signal rather than a complete site assessment.
Deep
- Original paper authors
- Abdullah Rehman, Quentin Paletta, Rui Song
- Submitted to arXiv
- License
- Not specified by source