Our Take on Meta’s New Maps

A recent publication by Meta and the World Resources Institute (WRI) promises an open-source global map of tree canopy height at a 1-meter resolution, allowing the detection of single trees at a global scale, a level of detail not seen before.

In the following post, we conclude that, although the researchers have conducted an impressive inventory of global forest heights, there are serious limitations for real world applications. We conducted a high-level sanity check for the possible utilization of the Meta canopy height product, specifically within the context of Chaco Vivo - the largest REDD+ project in Paraguay covering over 187,000 hectares of contiguous forest in the “Gran Chaco”, the second largest ecoregion in South America.

To compare the accuracy of the Meta product, we chose a publicly available dataset published by Lang et al in 2023, a 10-meter global map of canopy heights including the respective standard deviations of their modeling approach. We tested Meta's claim that their product improves vertical and horizontal accuracy of canopy height estimations.

Based on our investigation into these two datasets, we favor the approach of Lang et al. for the following three reasons:

• With the Meta dataset, we identify problematic artifacts in the form of stripes. At a very high zoom level, these stripes are not evident, however given the large area of the Chaco Vivo project at above 1,870 km² these stripes add a significant variability to our mapping and question the trustworthiness of the data. It's worth noting that the striping exists on different scales, with one general longitudinal gradient and many smaller stripes a few kilometers in width. See below for a direct comparison of canopy heights predicted by the two studies (same scaling).

• The Meta dataset lacks a predictive standard deviation layer as provided by Lang et al., which offers a globally consistent quality metric. While it is extremely difficult to create a reliable accuracy metric on the scale of “global forest heights”, this layer at least gives an indication of the quality of the produced maps. For our approach we are building capabilities to integrate multiple data sources and multi-sensor earth observation data. In a sense, this requires data sources to be harmonic and “team players”. By itself, a fusion product derived from GEDI, aerial LiDAR, and VHR Maxar imagery might be extremely powerful. Yet, while Meta’s product might be the overachieving star player in this ensemble approach, its lack of self awareness makes the dataset hard to deal with.

• Lastly but most importantly, reproducibility is highly relevant to our use case. Although the code, network, and logic are all accessible in a peer-reviewed paper in a prestigious journal, Meta’s approach is dependent on VHR imagery (WorldView and Quickbird) which is largely unattainable. The authors of the Meta paper highlight domain adaptation, yet it’s shown on exclusive aerial imagery as opposed to freely available Landsat/Sentinel-2. The approach is hence not applicable to monitoring or historical trend analysis, two key applications within REDD+.

At TransparenC we are on a mission to condense information of large-scale carbon offset projects, making it accessible and interpretable to all. We depend on high-quality data inputs and meta information to achieve our intended outcome. Although the study published by Meta and WRI is impressive within a technical and computational context, for applications that rely on multiple data sources and target real global transferability, we favor the more application-oriented approach by Lang et al.