This repository contains preprocessed data from the paper CROMA: Remote Sensing Representations with Contrastive Radar-Optical Masked Autoencoders published at NeurIPS 2023.

We did not create these datasets—if you use them, please cite the original papers!

All data is stored as PyTorch tensors (inside python dictionaries); images are normalized 8-bit integers. To use these data with CROMA, convert tensors to floats and divide by 255.

DFC 2020

dfc = torch.load("DFC_preprocessed.pt")  # pixel annotations with 8 classes
train_images = dfc["train_images"]  # shape (46_152, 14, 96, 96), first 12 channels are S2, last 2 channels are S1
train_labels = dfc["train_labels"]  # shape (46_152, 96, 96)
validation_images = dfc["validation_images"]  # shape (8_874, 14, 96, 96), first 12 channels are S2, last 2 channels are S1
validation_labels = dfc["validation_labels"]  # shape (8_874, 96, 96)

Naoto Yokoya, Pedram Ghamisi, Ronny Hänsch, and Michael Schmitt.
2020 IEEE GRSS Data Fusion Contest: Global Land Cover Mapping With Weak Supervision.
IEEE Geoscience and Remote Sensing Magazine, 2020.

Dynamic World (DW) — Expert

dw = torch.load("DynamicWorld_Expert_preprocessed.pt")  # pixel annotations with 9 classes (-1 as not-labeled)
train_images = dw["train_images"]  # shape (20_422, 14, 96, 96), first 12 channels are S2, last 2 channels are S1
train_labels = dw["train_labels"]  # shape (20_422, 96, 96)
validation_images = dw["val_images"]  # shape (51_022, 14, 96, 96), first 12 channels are S2, last 2 channels are S1
validation_labels = dw["val_labels"]  # shape (51_022, 96, 96)

Christopher F Brown, Steven P Brumby, Brookie Guzder-Williams, Tanya Birch, Samantha Brooks Hyde,
Joseph Mazzariello, Wanda Czerwinski, Valerie J Pasquarella, Robert Haertel, Simon Ilyushchenko, et al.
Dynamic World, Near real-time global 10 m land use land cover mapping.
Scientific Data, 2022.

MARIDA

marida = torch.load("MARIDA_preprocessed.pt")  # pixel annotations with 15 classes (-1 as not-labeled)
train_images = marida["train_images"]  # shape (1_682, 11, 96, 96)
train_labels = marida["train_labels"]  # shape (1_682, 96, 96)
validation_images = marida["validation_images"]  # shape (1_615, 11, 96, 96)
validation_labels = marida["validation_labels"]  # shape (1_615, 96, 96)

Katerina Kikaki, Ioannis Kakogeorgiou, Paraskevi Mikeli, Dionysios E Raitsos, and Konstantinos Karantzalos.
MARIDA: A benchmark for Marine Debris detection from Sentinel-2 remote sensing data.
PloS one, 2022.

Canadian Cropland

crops = torch.load("Canadian_Cropland_preprocessed.pt")  # image annotations with 10 classes

# 2017
train_images_2017 = crops["2017"]["train_imgs"]  # shape (9_898, 13, 65, 65)
train_labels_2017 = crops["2017"]["train_labels"]  # shape (9_898)
validation_images_2017 = crops["2017"]["val_imgs"]  # shape (2_075, 13, 65, 65)
validation_labels_2017 = crops["2017"]["val_labels"]  # shape (2_075)
test_images_2017 = crops["2017"]["test_imgs"]  # shape (2_138, 13, 65, 65)
test_labels_2017 = crops["2017"]["test_labels"]  # shape (2_138)

# 2018
train_images_2018 = crops["2018"]["train_imgs"]  # shape (12_789, 13, 65, 65)
train_labels_2018 = crops["2018"]["train_labels"]  # shape (12_789)
validation_images_2018 = crops["2018"]["val_imgs"]  # shape (2_714, 13, 65, 65)
validation_labels_2018 = crops["2018"]["val_labels"]  # shape (2_714)
test_images_2018 = crops["2018"]["test_imgs"]  # shape (2_822, 13, 65, 65)
test_labels_2018 = crops["2018"]["test_labels"]  # shape (2_822)

# 2019a
train_images_2019a = crops["2019a"]["train_imgs"]  # shape (11_628, 12, 65, 65)
train_labels_2019a = crops["2019a"]["train_labels"]  # shape (11_628)
validation_images_2019a = crops["2019a"]["val_imgs"]  # shape (2_486, 12, 65, 65)
validation_labels_2019a = crops["2019a"]["val_labels"]  # shape (2_486)
test_images_2019a = crops["2019a"]["test_imgs"]  # shape (2_517, 12, 65, 65)
test_labels_2019a = crops["2019a"]["test_labels"]  # shape (2_517)

# 2019b
train_images_2019b = crops["2019b"]["train_imgs"]  # shape (4_094, 13, 65, 65)
train_labels_2019b = crops["2019b"]["train_labels"]  # shape (4_094)
validation_images_2019b = crops["2019b"]["val_imgs"]  # shape (821, 13, 65, 65)
validation_labels_2019b = crops["2019b"]["val_labels"]  # shape (821)
test_images_2019b = crops["2019b"]["test_imgs"]  # shape (858, 13, 65, 65)
test_labels_2019b = crops["2019b"]["test_labels"]  # shape (858)

# 2020
train_images_2020 = crops["2020"]["train_imgs"]  # shape (15_475, 12, 65, 65)
train_labels_2020 = crops["2020"]["train_labels"]  # shape (15_475)
validation_images_2020 = crops["2020"]["val_imgs"]  # shape (3_318, 12, 65, 65)
validation_labels_2020 = crops["2020"]["val_labels"]  # shape (3_318)
test_images_2020 = crops["2020"]["test_imgs"]  # shape (3_339, 12, 65, 65)
test_labels_2020 = crops["2020"]["test_labels"]  # shape (3_339)

Amanda A Boatswain Jacques, Abdoulaye Baniré Diallo, and Etienne Lord.
Towards the Creation of a Canadian Land-Use Dataset for Agricultural Land Classification.
In 42nd Canadian Symposium on Remote Sensing: Understanding Our World: Remote Sensing for a Sustainable Future, 2021.