Characterizing Brazilian Atlantic Forest Restoration Outcomes with Geospatial AlphaEarth Embeddings
Alice Heiman · 2026
Read Characterizing Brazilian Atlantic Forest Restoration Outcomes with Geospatial AlphaEarth Embeddings by Alice Heiman (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-05547.txt.
Introduction
Brazil's Atlantic Forest is a biodiversity hotspot where only a small share of the original cover survives. Alice Heiman asks whether satellite-based AI embeddings can track how well cleared land is being restored. Ground reporting cannot scale to thousands of sites, and common vegetation indices saturate before young forest matures.
Heiman studied over seventeen hundred restoration sites in São Paulo using AlphaEarth embeddings. She defines success by similarity to mature secondary forest, a reference trajectory in embedding space. Sites clustered clearly by land use, and some showed clean change signals, though the embeddings still miss some site details.
Catalog record
| CREATOR | Alice Heiman |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 13 KB |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.05547
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Characterizing Brazilian Atlantic Forest Restoration Outcomes with Geospatial AlphaEarth Embeddings Alice Heiman Abstract The Atlantic Forest in Brazil is a critical biodiversity hotspot, yet less than 12–15% of its original cover remains. Although monitoring forest restoration on a large scale is essential, traditional methods are limited by the impracticality of on-the-ground reporting on such a scale and by the saturation of remote-sensing indices such as NDVI. Furthermore, reforestation is a gradual process as opposed to the rapid spectral changes caused by deforestation. In this study, we examine 1,729 restoration sites in São Paulo, using satellite embeddings from the AlphaEarth Foundation’s model to evaluate their effectiveness in characterising early restoration success. We introduce the concept of a “Reference Trajectory Embedding”, defining a metric of restoration success based on cosine similarity to reference sites of mature secondary forest. We observe distinct clusters in embedding space according to different land use and land cover (LULC) types, and we can identify sites with clear change vectors. However, the signal can be noisy, and embeddings may require further fine-tuning to capture and predict site metadata beyond LULC. 1 Introduction Brazil, short for ’Terra do Brasil’ (Eng. ’Land of Brazilwood’), is home to some of the world’s most important and biodiverse forests. However, the tree that gave Brazil its name is not found in the Amazon Rainforest. The endangered paubrasila tree can only be found in the Atlantic Forest. This forest provides ecosystem services to over 150 million Brazilians, contains over 20,000 plant species and supplies 62% of Brazil’s hydroelectric power ( 1 ; 4 ) . However, less than 12–15% of the original forest area remains, which is why it has been designated a World Restoration Flagship Project by the United Nations (UN) ( 3 ) . Initiatives such as the Trinational Atlantic Forest Pact have attracted over 300 organisations with the aim of restoring the forest. Yet, planting alone is insufficient. Effective restoration requires rigorous monitoring to understand what drives ecological success, but traditional vegetation monitoring metrics, such as NDVI, often saturate in dense tropical regions and lack universal thresholds for success ( 2 ) . Furthermore, remote sensing methods have traditionally focused on deforestation due to the clearer change in the spectral signal, rather than the gradual changes that occur in recovering forests over decades. Earth observation foundation models (FMs), such as AlphaEarth Foundations, produce learned embeddings that combine multiple data types. These high-dimensional vectors may encode more ecological information than simple indices. In this paper, we propose using geospatial embeddings to characterise reforestation sites. Specifically, we explore whether geospatial embeddings can be used to: (1) Track the early success of restoration projects in tropical settings. (2) Cluster restoration sites in ecologically meaningful ways based on embeddings; and (3) Infer restoration metadata better than features like NDVI, topology, and climate alone. Our contributions include a novel “Reference Trajectory Embedding” success metric that uses cosine similarity to mature secondary forest reference sites, as well as a systematic evaluation of 1,729 reforestation polygons in the Atlantic Forest and their restoration trajectories in embedding space. 2 Data 2.1 Study Sites We select a subset of 43,139 restoration polygons from the Observatorio da Restauração e Reflorestamento (ORR). Specifically, we filter for Atlantic Forest sites in the state of São Paulo ( n=14,873 ) with a restoration start year between 2017 and 2024 ( n=14,526 ). We only use sites with an area of at least 1 hectare after applying a -10 m buffer to avoid edge effects ( n=1,729 ). We use São Paulo as a case study because it contains the most complete metadata. We limit the analysis to the period 2017–2024, as this was the availability of the pre-computed AlphaEarth embeddings. Figure 1: Study area in the Atlantic Forest biome, Brazil, with restoration site polygons and reference LULC points marked. 2.2 Reference Point Sampling We use the MapBiomas Brazil Collection 10 Coverage V1 and Deforestation Secondary Vegetation V2 layers in Google Earth Engine to sample two categories of reference points. Stable points share the same land use and land cover (LULC) class for 10 or more consecutive years through 2024; changing points transition from one LULC class (2017-2020) to another (2021-2024). We sample n{=}1{,}000 points per class for nine stable classes: Primary Forest, Secondary Forest, Forest Plantation, Wetland, Sugar Cane, Coffee, Grassland, Pasture, and Urban. We sample n{=}200 points per transition for six changing classes: Forest Formation \rightarrow Urban, Forest Formation \rightarrow Sugar Cane, Forest Formation \rightarrow Pasture, Pasture \rightarrow Forest Formation, and Coffee \rightarrow Forest Formation. 2.3 Data Extraction We extract the following data at each reference point or as a restoration polygon mean: AlphaEarth Embeddings (64 dimensions per site per year); Vegetation Indices (monthly and annual Sentinel-2-derived NDVI and EVI); Environmental Covariates (annual precipitation, min/max temperature, elevation, slope, aspect, evapotranspiration, forest cover within 2 km, and road density within 5 km); and Project Metadata (restoration area, start year, and strategy from the ORR dataset). See Appendix A.1 for a detailed breakdown of data sources used. 3 Methods and Results 3.1 Embedding Space Visualization To verify the semantics of the AlphaEarth embedding space, we fit a UMAP on all stable LULC reference point embeddings from 2024 and visualize them in 2D space. As shown in Figure 2 , the LULC classes form distinct clusters with interpretable structure: the Urban class is the most separated, while more heterogeneous classes such as Forest exhibit greater intra-class spread. Notably, we were able to reveal some misclassifications in the original MapBiomas LULC maps, where points labeled as one class cluster closer to a different class in embedding space (see Appendix A.5 ). We further visualize the year-on-year embedding trajectories of known changing reference points to qualitatively verify that transitions between LULC classes are reflected in the embedding space. Figure 2: AlphaEarth embeddings of reference stable and changing points projected into 2D using UMAP. 3.2 Restoration Trajectories We compute a global reference vector \bar{\mathbf{r}} as the mean embedding of all sampled stable secondary forest pixels. The global similarity of polygon i at year t is s^{\text{global}}_{i,t}=\frac{\mathbf{e}_{i,t}\cdot\bar{\mathbf{r}}}{\|\mathbf{e}_{i,t}\|\,\|\bar{\mathbf{r}}\|}, (1) where \mathbf{e}_{i,t}\in\mathbb{R}^{64} is the polygon’s embedding at year t . To control for regional variation in forest composition, we also compute a local reference for each polygon. To do so, we identify the embedding of the closest reference secondary forest point \bar{\mathbf{r}}_{i} to each restoration polygon. The local similarity s^{\text{local}}_{i,t} is then the cosine similarity between \mathbf{e}_{i,t} and \bar{\mathbf{r}}_{i} . Each polygon thus has a time series \{s_{i,t}\}_{t=0}^{T_{i}} , where t=0 corresponds to the restoration start year. To control for different start years, plots show the \Delta t , i.e. the years since restoration start year. Thus, the similarity scores s_{i,t} can be interpreted as absolute restoration progress scores and visualized as year-on-year trajectories. The difference, meanwhile, can be interpreted as a relative improvement score from the start of the restoration process: s_{i,T_{i}}-s_{i,0} . We use stable-class means (primary forest and pasture) to provide upper and lower similarity baseline references to contextualise the trajectories. We also compare the trajectories to the NDVI and EVI spectral indices (see Figure 3 ). Figure 3: Similarity, NDVI, and EVI trajectories stratified by starting MapBiomas LULC class. 4 Prediction Tasks We pose two tasks for predicting