Checkerboard Dataset: hyperspectral unmixing dataset with ground truth

Description

This dataset contains seven hyperspectral images of a printed checkerboard pattern, specifically designed for research into mixed pixels and spectral unmixing in the presence of spectral variability. Each hyperspectral image was acquired using an Imec Snapscan VNIR camera under varying illumination conditions (different incident light angles and intensities). This controlled variability in lighting was introduced to generate spectral diversity within the scenes, mimicking real-world complexities.

The checkerboard pattern itself is composed of four distinct colors: black, yellow, red, and blue. It is designed to include all possible mixtures of these four colors at 25% intervals, allowing for the generation of mixed pixels with known abundance values.

The primary purpose of this dataset is to provide a controlled environment for studying mixed pixels that exhibit spectral variability. The varying illumination conditions applied during image acquisition lead to changes in the spectral signatures of the pure endmembers across the scenes, which is crucial for developing and testing robust unmixing algorithms. The known abundance values of the mixed pixels (due to the checkerboard design) provide ground truth for quantitative analysis.

Structure

The dataset is organized into the following components:

images/: This folder contains:

.mat files, each representing a hyperspectral image of the checkerboard (dimensions: 400x400 pixels, 150 spectral bands)

Matlab scripts designed to extract mixed and pure pixels from these images. Each image yields 42 distinct mixed and pure pixel samples.

whitepanels/: This folder contains:

.mat files corresponding to scans of a diffuse white panel. These scans can be used to empirically determine the intensity of a given scene and to assess the degree of sensor-induced spectral variability.

load_checkerboard.m: This is a Matlab function that facilitates the extraction of the entire dataset from the raw data. It offers several configurable options. Refer to this function's documentation for more details.

data.mat: This file contains a Matlab struct, which is the result of running the load_checkerboard.m function with all its default options. This provides a pre-processed version of the dataset for quick access and initial exploration.

Citation

If you use this dataset in your research or work, please cite it using the corresponding DOI: doi.org/10.5281/zenodo.15632215

Changes

Version 2 corrects small bugs in the MATLAB scripts, more specifically a bug in the extract_squares scripts which led to wrong ground truth abundances for pure pixels.