Vision Transformers for High-Resolution Canopy Height Mapping

tbd

Available Models

Model	Architecture	Params
vit-nano	Vision Transformer + ConvNet	3.049.665
vit-micro	Vision Transformer + ConvNet	6.576.777
vit-tiny	Vision Transformer + ConvNet	14.518.713
vit-small	Vision Transformer + ConvNet	25.563.369
unet	U-Net	10.103.933
unet++	U-Net++	10.479.187

Training Data

• 2693 multispectral PlanetScope satellite images (4096x4096x4, ~650 GB)
• 4 channels per image (R, G, B, NIR)
• 505.724 patches (256x256) after preprocessing
• 16.933.743 GEDI labels after preprocessing

Install

To install the required packages to a virtual environment and activate it, run:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.sample .env

For merging the output files into one, you need to install GDAL:

brew install gdal

Usage

python [preprocessing|train|evaluate|run].py

Inference

python run.py

cd output

ls *.tif > merge_list.txt

gdal_merge.py -o output.tif --optfile merge_list.txt -co COMPRESS=LZW -co BIGTIFF=YES -n 0 -a_nodata 0

References

1. Vision Transformers for Dense Prediction, DPT

2. The overlooked contribution of trees outside forests to tree cover and woody biomass across Europe, planet_canopy_height_v0