Hyperspectral imagery for material surface damage

We provide a unique dataset with hyperspectral images (hyperspectral cubes) of 50x50x139 in size. Each cube is companied with a 1000x1000 high-resolution gray-level image, and the ground-truth mask. The pixel location mapping is straightforward. Each hyperspectral pixel corresponds to a 20x20 neighborhood pixels in the gray image (and in the mask). In the mask image, color is used to indicate the class labels. Besides the underlying background pixels (which are either concrete or asphalt as contained in different folders), six different classes are labeled and rendered in color in the mask images. They include cracking, green vegetation, dry vegetation, water, oil, and artificial marking, which are assigned with the color of black, green, brown, blue, red, and yellow, respectively.
We provide this dataset and encourage freely academic research and educational use. We expect that our dataset can be used to advance the research in machine learning (deep learning), dimensionality reduction, feature extraction, non-/semi-supervised learning, material damage detection, and civil infrastructure assessment in general.