OxHyperSyntheticCH4 dataset

Author: Vít Růžička

One of the OxHyper datasets, for more details please check my main web at: https://previtus.github.io/

Fast data preview in: https://huggingface.co/datasets/previtus/OxHyperSyntheticCH4/blob/main/dataset_exploration.ipynb

The "OxHyperSyntheticCH4" dataset is a bi-temporal hyperspectral dataset of synthetically simulated methane leak events using the labels of the STARCOP dataset (https://github.com/spaceml-org/STARCOP) with newly downloaded near-global hyperspectral datacubes from the EMIT sensor. In total we keep 796 large tiles (of 512x512 px) in the training dataset, 198 in the validation dataset and 200 in the test dataset. Each set is made from completely non-overlapping sources of data (clean EMIT datacubes and labels from the STARCOP dataset from AVIRIS-NG data). In each subset we uniformly sample the available plume sizes, so that we have similar distribution of event sizes. Additionally, half of these tiles contain simulated events. We note that these large tiles (of 512x512 px) are further tiled when training the models. When using the tile size of 64 pixels with an overlap of 32 pixels, we finally get a training dataset of 174341 samples. The total size of this data is 228 GB. It contains 86 spectral bands made of the RGB bands and ranges between 1573-1699nm and 2004-2478nm.

More info will be added upon publication, for now please refer to https://previtus.github.io/

More information about the NASA's sensor EMIT: https://earth.jpl.nasa.gov/emit/

Variants

The OxHyperSyntheticCH4 dataset has three variants depending on the number of included bands.

mono-temporal, 86 bands (~114GB) https://huggingface.co/datasets/previtus/OxHyperSyntheticCH4/tree/main

multi-temporal, 86 bands (~228GB) https://huggingface.co/datasets/previtus/OxHyperSyntheticCH4/tree/ab_multitemporal

RGB + MF only (~6GB) https://huggingface.co/datasets/previtus/OxHyperSyntheticCH4/tree/b_rgb_mf

Please note that we will keep updated readme only on the main branch

File structure

Each folder represent a large tile (of 512x512px). Rows in the attached csv's split these into train/val/test subsets.

Each folder contains these files:

• B - all 86 bands in a georeferenced file (note: you'll also need the B.hdr file to load it, can be also inspected for example in QGIS)
• B_EMIT_641nm.tif, B_EMIT_551nm.tif, B_EMIT_462nm.tif - The RGB bands, these were extracted from the larger file
• labelbinary.tif - binary label (pixel values 0 or 1 for "no plume" or "plume" class)
• B_magic30_tile.tif - computed matched filter product (using the iterative variant called mag1c)
• info.json - saved information about the used source data (raw EMIT file, the AVIRIS-NG file used for methane leak simulation and also a flag stating if this large tile contains a plume leak event)
• (in multi-temporal case) A and corresponding extracted files are the datacube for earlier observation (these also don't contain any methane leak events)
• Note: for each georeferenced tif we also provide the .hdr file which is needed for loading and has details about the bands

Citation

Will be added.