sfoucher / seasonal-contrast

Seasonal Contrast: Unsupervised Pre-Training from Uncurated Remote Sensing Data

This is the official PyTorch implementation of the SeCo paper:

@article{manas2021seasonal,
  title={Seasonal Contrast: Unsupervised Pre-Training from Uncurated Remote Sensing Data},
  author={Ma{\~n}as, Oscar and Lacoste, Alexandre and Giro-i-Nieto, Xavier and Vazquez, David and Rodriguez, Pau},
  journal={arXiv preprint arXiv:2103.16607},
  year={2021}
}

Preparation

Install Python dependencies by running:

pip install -r requirements.txt

Data Collection

First, obtain Earth Engine authentication credentials by following the installation instructions.

Then, to collect and download a new SeCo dataset from a random set of Earth locations, run:

python datasets/seco_downloader.py \
  --save_path [folder where data will be downloaded] \
  --num_locations 200000

Unsupervised Pre-training

To do unsupervised pre-training of a ResNet-18 model on the SeCo dataset, run:

python main_pretrain.py \
  --data_dir datasets/seco_1m --data_mode seco \
  --base_encoder resnet18

Transferring to Downstream Tasks

With a pre-trained SeCo model, to train a supervised linear classifier on 10% of the BigEarthNet training set in a 4-GPU machine, run:

python main_bigearthnet.py \
  --gpus 4 --accelerator dp --batch_size 1024 \
  --data_dir datasets/bigearthnet --train_frac 0.1 \
  --backbone_type pretrain --ckpt_path checkpoints/seco_resnet18_1m.ckpt \
  --freeze_backbone --learning_rate 1e-3

To train a supervised linear classifier on EuroSAT from a pre-trained SeCo model, run:

python main_eurosat.py \
  --data_dir datasets/eurosat \
  --backbone_type pretrain --ckpt_path checkpoints/seco_resnet18_1m.ckpt

To train a supervised change detection model on OSCD from a pre-trained SeCo model, run:

python main_oscd.py \
  --data_dir datasets/oscd \
  --backbone_type pretrain --ckpt_path checkpoints/seco_resnet18_1m.ckpt

Datasets

Our collected SeCo datasets can be downloaded as following:

#images	RGB preview	size	link	md5
100K	✓	7.3 GB	download	ebf2d5e03adc6e657f9a69a20ad863e0
~1M		36.3 GB	download	187963d852d4d3ce6637743ec3a4bd9e

Pre-trained Models

Our pre-trained SeCo models can be downloaded as following:

dataset	architecture	link	md5
SeCo-100K	ResNet-18	download	dcf336be31f6c6b0e77dcb6cc958fca8
SeCo-1M	ResNet-18	download	53d5c41d0f479bdfd31d6746ad4126db
SeCo-100K	ResNet-50	download	9672c303f6334ef816494c13b9d05753
SeCo-1M	ResNet-50	download	7b09c54aed33c0c988b425c54f4ef948