Dataset Card for BioMassters for Global Prithvi

Dataset Description

Copied from BioMassters: A Benchmark Dataset for Forest Biomass Estimation using Multi-modal Satellite Time-series https://nascetti-a.github.io/BioMasster/

• Original Dataset: https://huggingface.co/datasets/nascetti-a/BioMassters
• Point of Contact For Updated Dataset: Denys Godwin (dgodwin@clarku.edu)

Dataset Summary

This dataset contains Sentinel-1 SAR and Sentinel-2 MSI imagery of Finnish forests for the years 2016-2021. There are 11,462 reference images of Above Ground Biomass (AGB). Each reference AGB has corresponding S1 and S2 imagery for the 12 months leading up to the AGB observation. Sentinel-1 data exists for each of the 12 months, while Sentinel-2 data does not have full temporal coverage. All imagery has height and width of 256x256.

Modifications for Prithvi

The directory structure of the dataset has been preserved. However, to enable the filtering needed for Prithvi training, some chip metadata has been calculated and stored in biomassters_chip_tracker.csv. Additional metadata are as follows:

• cloud_percentage: the percentage of clouds, defined as pixels where the cloud probability band of the input image exceeds 70
• corrupt_values: boolean representing whether there are corrupt values in the image. These are defined as blocks of identical values, which are present in some images
• red_mean: the mean value of the red band (band 3) of the Sentinel-2 image, used to filter scenes with anomalously high reflectance due to snow and Top of Atmosphere (ToA) correction issues

Data Splits

Training and testing splits are as given by The BioMassters original dataset. The training dataset has been further randomly split into training (80%) and validation (20%) sets. These are tracked in the 'split' column of the biomassters_chip_tracker.csv

Reference data:

• Reference AGB measurements were collected using LiDAR (Light Detection and Ranging) calibrated with in-situ measurements.
• Total 11,462 patches, each patch covering 2,560 by 2,560 meter area.

Feature data:

• Sentinel-1 SAR and Sentinel-2 MSI data
• 12 months of data (1 image per month)
• Sentinel-2 data does not exist for all months for all scenes
• Sentinel-2 data contains 10 selected spectral bands and 1 band of cloud probability

Dataset Specification

Data Size:

  dataset         | # files  | size
  --------------------------------------
  train_features  |  189078  | 215.9GB
  test_features   |   63348  |  73.0GB
  train_agbm      |    8689  |   2.1GB
  test_agbm       |    2773  |   705MB

Citation : under review