Authors: Ghjulia Sialelli ([email protected]), Torben Peters, Jan Wegner, Konrad Schindler
This repository contains the code used to create the results presented in the eponymous paper. We curated a dataset from various remote-sensing data sources (Sentinel-2 L2A, ALOS-2 PALSAR-2 yearly mocaics, JAXA Digital Elevation Model, Copernicus Land Cover, Lang et al. Canopy Height Map) and GEDI L4A Above-Ground Biomass (AGB) data. We developed benchmark models for the task of estimating Above-Ground Biomass (AGB).
To install the packages required to run this code, you can simply run the following commands, which will create a conda virtual environment called agbd. For more details, follow the instructions on pytorch.org.
conda create -n agbd python=3.10.9 pytorch torchvision torchaudio cudatoolkit=11.8 -c pytorch -c nvidia
conda env update -n agbd -f requirements.yml
conda activate agbd
conda create -n agbd python=3.10.9 pytorch::pytorch torchvision torchaudio -c pytorch
conda env update -n agbd -f requirements_mac.yml
conda activate agbd
The dataset is openly accessible on HuggingFace, where it is stored in a streamable ML-ready format. You can use it as follows:
#!pip install datasets
from datasets import load_dataset
dataset = load_dataset('prs-eth/AGBD', trust_remote_code=True, streaming=True)["train"] # Options: "train", "val", "test"
Should you wish to reproduce our results, we provide in the Models section of this repository the code we used to train our benchmark models. Should you want to reproduce our experiments with the data format we used, you can download the data and use the provided data loader. You can find further instruction on how to do it in the dedicated folder. Note that the data is ~300GB.
We provide an example for the patches creation procedure, in the Patches section of this repository. Further instructions can be found there.
Our dense predictions for the region covered by the dataset can be downloaded via the following command (it represents ~40GB) :
wget "https://libdrive.ethz.ch/index.php/s/VPio6i5UlXTgir0/download?path=%2F&files=predictions&downloadStartSecret=gxairgqzc" -O predictions.tar
You will get a .tif file per Sentinel-2 tile in the regions of interest.
Please note that those dense predictions pre-date the "latitude / longitude bug" (see the changelog for more information). As we are currently working on a better model, we do not generate the post-bug predictions, but will directly upload the best ones shortly.
See the changelog for more information about what was updated with each new commit (when relevant).
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
