This interactive map displays tree canopy heights derived from aerial imagery via an AI model trained by Meta and the World Resource Institute on about 18 million high resolution satellite images from around the globe.
AI, deep learning and transformer architectures are revolutionizing geospatial remote sensing analytics. The recent breakthroughs in natural language processing for model pretraining on large quantities of data have opened the way for similar foundation models in geospatial remote sensing.
For details of the Meta approach, see the paper HERE.
ThermaInsights forked the Meta GitHub and configured the aerial inference model to estimate tree canopy height from NAIP imagery collected over municipalities in the USA. The layer shown is canopy height from NAIP over Boulder, CO.
Interested in tree canopy height for your city? Contact Us for a free sample.
This map shows the distribution of tree cover mapped using semantic segmentation with aerial NAIP imagery. DetecTree is a Pythonic library to perform semantic segmentation of aerial imagery into tree/non-tree pixels, following the methods of Yang et al. A pre-trained model is available at Hugging Face.
Sample code for running the model can be accessed via GitHub, below.
This example, when compared to the canopy height output, serves to demonstrate the leep in capability afforded over the last half-decade by new vision transformer architectures, and very large models trained on large quantities of data.
Check out the code on GitHub:
View Repository