Skip to content

Latest commit

 

History

History
103 lines (67 loc) · 3.4 KB

README.md

File metadata and controls

103 lines (67 loc) · 3.4 KB

Downstream tasks

We provide scripts for reproducing our the pretraining of our SeCo-Eco and MoCo-Eco models on our SSL4Eco dataset.

Note: If you have not done it already, make sure you downloaded the desired downstream datasets. See the data download section for this.

Note: If you intend to use our pretrained SeCo-Eco model, make sure you downloaded the weights from huggingface 🤗.

We use hydra for managing our script's arguments here. You can find the default config in default.yaml. The syntax for running the downstream evaluation of a model is simple:

# Set the SECOECO_ROOT variable to indicate where the datasets can be 
# found and where the outputs will be saved
export SECOECO_ROOT=/root/path/to/your/datasets

# Evaluate SeCo-Eco on the biomes dataset, with linear probing
python main.py MODEL=secoeco DATASET=biomes PROBE=linear

# Evaluate Satlas on the arctic dataset, with k-NN
python main.py MODEL=satlas DATASET=arctic PROBE=knn

The complete list of supported models can be found here:

'ablcalendar'
'croma'
'dofabase'
'dofalarge'
'mocoeco'
'satlas'
'satmae'
'seco'
'secoeco'
'ssl4eo'

The complete list of supported datasets can be found here:

'arctic'
'bigearthnet'
'biomassters'
'biomassterswinter'
'biomes'
'clef'
'clefblind'
'climate'
'euforest'
'euforestsp'
'treesatai'

Downstream tasks preparation

Biomes, Arctic, Climate, EUForest, EUForestSp

Please follow the instructions of data download section for image download. The corresponding index iles are in the index_files folder. The datasets are ready to use.

BigEarthNet 10%

To prepare the BigEarthNet 10% dataset, please download the images from here and place the BigEarthNet-v1.0 folder into the BE_full folder.

Copy the train-test split into the BE_full folder:

cp ./index_files/be10/bigearthnet-train.csv ./BE_full/bigearthnet-train.csv
cp ./index_files/be10/bigearthnet-test.csv ./BE_full/bigearthnet-test.csv

To add NDVI band to BigEarthNet imagery, run

python ./downstream_tasks/add_ndvi_band_BE.py ./BE_full/imgs

CLEF, CLEF blind

The images for CLEF should be downloaded similar to other datasets to the folders sdm/clef/imgs based on coordinates from /index_files/sdm/clef_labels.csv index file. The blind part of the competition should be downloaded into sdm/clef_blind/imgs based on coordinates from /index_files/sdm/test_blind.csv index file.

The final score can be obtained by submitting the resulting prediction into the official Leaderboard.

BioMassters

Please download original dataset from here. The original data contain all Sentinel-2 bands except 1, 9 and 10. We only drop band 10, therefore band 1 and 9 need to be added manually as zero matrices along with the NDVI-band through the pre-proccessing script. From this point on, the dataset code automatically takes care of parsing the images and turning the pixel-wise AGBM labels into a 1-D label vector using binning.