Amandeep10/pretrain-where-pretraining-datasets

Dataset Card for 'Pretrain-where' pretraining datasets (SatMAE)

These datasets were used to pretrain SatMAE to study the impact of spatial context of a pretraining dataset on downstream performance (Main paper: https://arxiv.org/abs/2604.21104).

Dataset Details

Dataset Description

The collection consists of 7 pretraining datasets, each with data from a specific spatial context, namely Africa only (af), Asia only (as), North America only (na), South America only (sa), Europe only (eu), Oceania only (oc), and global(gl). Each dataset consists of 700k samples of Sentinel-2 imagery. Each image is ~96x96 and pre-normalised.

• Curated by: Amandeep Kaur

Dataset Sources

Data consists of Sentinel-2 imagery, originally downloaded using Microsoft Planetary computer (https://planetarycomputer.microsoft.com/dataset/sentinel-2-l2a).

Uses

These datasets are intended to be used for pretraining Geospatial Foundational Models (GeoFMs). Since, pretrained SatMAE weights will also be made public, a novel use could be to create new datasets by combining these datasets in different ratios and studying the impact of such a mixture.

Dataset Structure

The collection contains 7 pretraining datasets namely as (asia only), af (africa only), eu (Europe only), na (North America only), sa (South America only), oc (Oceania only) and gl (global). Each dataset folder contains two subfolders - train_shards (training data) and val_shards (validation data). The validation data is only used to monitor model behaviour as pretraining hyperparameters are held fixed for our work. A few datasets might not contain validation data, in such case, if validation is required, training data can be split to create some validation data. In our datasets, training data is ~700k samples and validation is ~20k.

Dataset Creation

Each dataset was created in 2 steps: 1) Sample datapoints (lat, lon) - Uniformaly at random using QGIS 2) Images were downloaded corresponding to sampled locations using Microsoft Planetary Compute. The downloaded images are preprocessed to decrease storage space, preprocessing and further sampling is defined in following sections.

Curation Rationale

The aim of the study was to measure the effect of changing the pretraining dataset while keeping all other variables constant (model architecture, pretraining hyperparameters, downstream tasks, preprocessing pipelines). In order to vary the pretraining dataset spatially, we choose continents as the base unit. We created one dataset per continent (leaving out Antartica) and one global dataset. We choose uniform at random (UAR) sampling as it's a good default as it's free of any assumptions (A sampling scheme biased towards maximising biomes/ecoregions or towards urbanisation of city centers assumes these samples of be of superior quality, but these are the trypes of assumptions we want to test). Rest of the decisions were based of the model, SatMAE.

Source Data

The source of the samples is Sentinel-2 satellite. Data is publicily available from Microsoft Planetary Compute and Google Earth Engine (throtled).

Data Collection and Processing

The downloaded samples were limited to a maximum of 20% cloud cover and the images were taken from the year 2024. As SatMAE only uses 10 bands in the channel-grouped setting, we only download the 10 bands, namely B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12. We also pre-normalise these bands using the following means and std values: S2_MEAN = [1184.3824625 , 1120.77120066, 1136.26026392,1263.73947144, 1645.40315151, 1846.87040806, 1762.59530783,1972.62420416, 1732.16362238, 1247.91870117] S2_STD = [650.2842772 , 712.12507725, 965.23119807,948.9819932 , 1108.06650639, 1258.36394548, 1233.1492281 ,1364.38688993, 1310.36996126, 1087.6020813] We save the values at uint16.

Bias, Risks, and Limitations

No bias, risks or limitations are noted by the authors. Note that some datasets might have fewer number of samples i.e. 650K instead of 700K which could cause change in model performance but it was not seen in our experiments.

Citation [optional]

Please cite the original paper when mentioning the datasets or our work. BibTeX:

@misc{kaur2026pretrainwhereinvestigatingpretraining, title={Pretrain Where? Investigating How Pretraining Data Diversity Impacts Geospatial Foundation Model Performance}, author={Amandeep Kaur and Mirali Purohit and Gedeon Muhawenayo and Esther Rolf and Hannah Kerner}, year={2026}, eprint={2604.21104}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2604.21104}, }

Dataset Card Authors

Amandeep Kaur, please reach out to me for any questions/concerns at akaur64@asu.edu