Tropical forests host a large share of the world’s biodiversity, regulate climate by storing carbon, and provide essential natural resources and ecosystem services both globally and to local communities. The Brazilian Atlantic Forest is one of the most biologically rich regions on Earth, with a high level of endemism, but today only a small and highly fragmented part of the original forest remains after centuries of human activity, primarily cattle farming. Restoring these forests would be hugely beneficial for preserving biodiversity, mitigating climate change, and improving human well-being. Given limited restoration resources, there is a need to prioritise and focus on areas where restoration would be the most successful and beneficial in terms of biodiversity and carbon storage, while incurring the lowest implementation and opportunity costs possible. To make sound decisions of spatial prioritisation, it is necessary to acquire both a clear understanding of current states of the forest and the ability to predict patterns of biodiversity and carbon storage recovery at broad scales, as the forests grow back through natural and human-mediated reforestation. Geospatial foundation model is a new powerful machine learning approach that integrates various sources of spatially explicit large datasets into harmonized embeddings that can be used in downstream tasks such as spatial prediction of biodiversity patterns and ecosystem functioning across broad scales. The aim of this project is to construct a foundation model that encodes biodiversity information and environmental characteristics of the Brazilian Atlantic Forest ecoregion, combining remote sensing and field data. In particular, we will generate high-resolution (30-m) species suitability maps for ~4000 vascular plant species occurring in the region through species distribution modelling (SDM), based on a curated species occurrence dataset. The species suitability maps will be used alongside geospatial data layers such as climate, soil, topography and historical land use as inputs for the construction of the foundation model, which can then be combined with field that we will either collect personally or obtain from collaborators to predict biodiversity change across a chronosequence of forest regeneration under various environmental conditions, to predict spatially the maximum benefits that forest restoration activities can bring in terms of biodiversity recovery, ecosystem functioning improvement, and carbon sequestration. This knowledge will allow us to plan restoration better by identifying areas where forest recovery delivers the greatest benefits.