This repository contains the dataset and metadata to support the double-blind review process. The dataset provides Sentinel-2 image patches from the Brazilian Legal Amazon sampled across different mining proportion ranges.
All images are provided in PNG format and are organized into folders based on their mining proportion interval.
0/: 1300 images with 0% mining0_10/: 130 images with 0% to 10% mining10_20/: 130 images with 10% to 20% mining20_30/: 130 images with 20% to 30% mining30_40/: 130 images with 30% to 40% mining40_50/: 130 images with 40% to 50% mining50_60/: 130 images with 50% to 60% mining60_70/: 130 images with 60% to 70% mining70_80/: 130 images with 70% to 80% mining80_90/: 130 images with 80% to 90% mining90_100/: 130 images with 90% to 100% mining
The manifest.csv file contains the detailed information for each image patch. The columns include the image identifier, the assigned mining proportion range, the geographic bounding box coordinates, and a suggested split for training and testing.
Due to the way Google Earth Engine exports spatial grids, the actual image dimensions vary slightly around the 128x128 pixel target size. We strongly recommend applying a standard center crop or resizing operation to unify the input shapes before feeding the patches into any machine learning model.