Awesome

CROMA

Official repository of CROMA: Remote Sensing Representations with Contrastive Radar-Optical Masked Autoencoders (NeurIPS '23)

Available on arxiv: https://arxiv.org/pdf/2311.00566.pdf

NeurIPS page: https://nips.cc/virtual/2023/poster/70928

Using Pretrained CROMA models

My goal is to make CROMA as noob-friendly as possible. To use pretrained CROMA models, you will need the use_croma.py file and pretrained weights. The pretrain_croma.py file is only needed if you want to pretrain your own models from scratch.

Download the pretrained weights here: https://huggingface.co/antofuller/CROMA/tree/main

Or using wget:

wget https://huggingface.co/antofuller/CROMA/resolve/main/CROMA_base.pt
wget https://huggingface.co/antofuller/CROMA/resolve/main/CROMA_large.pt

Install einops (any relatively recent version of einops and torch should work fine, please raise an issue otherwise)

pip install einops

Load Sentinel-1 & 2 images and preprocess them. CROMA's default image size is 120x120px, but this can be changed. Sentinel-1 must be 2 channels and Sentinel-2 must be 12 channels (remove the cirrus band if necessary).

import torch
from use_croma import PretrainedCROMA

device = 'cuda:0'  # use a GPU
use_8_bit = True
N = 1_000  # number of samples
sentinel_1 = torch.randn(N, 2, 120, 120).to(device)  # randomly generated for demonstration
sentinel_2 = torch.randn(N, 12, 120, 120).to(device)  # randomly generated for demonstration

def normalize(x, use_8_bit):
    # taken from SatMAE and SeCo
    x = x.float()

    imgs = []
    for channel in range(x.shape[1]):
        min_value = x[:, channel, :, :].mean() - 2 * x[:, channel, :, :].std()
        max_value = x[:, channel, :, :].mean() + 2 * x[:, channel, :, :].std()

        if use_8_bit:
            img = (x[:, channel, :, :] - min_value) / (max_value - min_value) * 255.0
            img = torch.clip(img, 0, 255).unsqueeze(dim=1).to(torch.uint8)
            imgs.append(img)
        else:
            img = (x[:, channel, :, :] - min_value) / (max_value - min_value)
            img = torch.clip(img, 0, 1).unsqueeze(dim=1)
            imgs.append(img)

    return torch.cat(imgs, dim=1)

sentinel_1 = normalize(sentinel_1, use_8_bit)
sentinel_2 = normalize(sentinel_2, use_8_bit)

PretrainedCROMA initializes the model(s) and loads the pretrained weights.

model = PretrainedCROMA(pretrained_path='CROMA_base.pt', size='base', modality='both', image_resolution=120).to(device)

if use_8_bit:
    sentinel_1 = sentinel_1.float() / 255
    sentinel_2 = sentinel_2.float() / 255

with torch.no_grad():
    outputs = model(SAR_images=sentinel_1, optical_images=sentinel_2)
"""
outputs is a dictionary with keys:
'SAR_encodings' --> output of the radar encoder, shape (batch_size, number_of_patches, dim)
'SAR_GAP' --> output of the radar FFN (after global average pooling (GAP)), shape (batch_size, dim)
'optical_encodings' --> output of the optical encoder, shape (batch_size, number_of_patches, dim)
'optical_GAP' --> output of the optical FFN (after global average pooling (GAP)), shape (batch_size, dim)
'joint_encodings' --> output of the joint radar-optical encoder, shape (batch_size, number_of_patches, dim)
'joint_GAP' --> global averaging pooling the joint_encodings, shape (batch_size, dim)
"""

Please Cite

@inproceedings{fuller2023croma,
  title={CROMA: Remote Sensing Representations with Contrastive Radar-Optical Masked Autoencoders},
  author={Fuller, Anthony and Millard, Koreen and Green, James R},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}

Preprocessed Benchmarks

We uploaded versions of publicly available benchmarks that we preprocessed—to help others build on CROMA. If you use them, please cite the original papers!!! https://huggingface.co/datasets/antofuller/CROMA_benchmarks