A dataset of 1600 images extracted from 5 cm RGB orthophotos for the classification of 12 classes of roofing materials

This dataset contains a collection of 1601 image tiles of 64x64 pixels (3.2x3.2m²) annotated for 12 roofing materials. These tiles were extracted from 5 cm RGB orthophotos acquired by the city of Namur (Belgium) in 2017. The additional data used to create this dataset are (a) a Namur roof section mask, and (b) a set of 1601 material samples acquired using stratified random sampling. The tiles were obtained as follows: the centroid of each roof section containing a sample is used to extract tiles. A size of 64x64 pixels has been chosen so that a tile contains information for only one roof section, in order to learn only the colour and texture of the roof materials. This also avoids adding information outside the given roof section. The tiles are thus extracted for each orthophoto spectral band and labelled with the identifier of the class of roofing materials to which they belong. Here are the 12 material classes considered, preceded by their labels:

0- Solar panels1- Brown tiles2- Orange tiles3- Black tiles4- Dark membranes5- White membranes6- Slates containing asbestos7- Slates without asbestos8- Corrugated asbestos-cement sheets9- Gravel10- Vegetation12- Metals

There are approximately 140 tiles by material class except for the vegetated roof sections (class 10) which contains only 47 samples due to its rarety.

The dataset contains 1 folder for each spectral band. Each folder contains 1601 thumbnails in tif format named as follows:

img[tile id]_[class label].tif

It is suggested to apply pre-processing to these images as done by Wyard et al. (2023).