Buildings are subject to environmental degradations therefore they must be maintained. The checking of these degradations is time consuming because everything is done by hand. In order to improve the efficiency, unmanned aircraft systems has already been coupled with hyper-spectral imagery (HSI) in [1]. In this project, we combine deep learning (DL) with HSI to classify different materials with semantic segmentation, as a first step to solve this problem. The hyper-spectral images come from a FX17 hyper-spectral camera. It works in the near infrared region (NIR), from 900 to 1700nm, with 224 bands. Four datasets were used during this project: a synthetic dataset created by hand, a plastics dataset from a prior project, one with tiles of painted woods and one containing images of the client’s building. Four DL models were used: U-Net, U-Net++, deep residual U-Net and DeepLabV3+ with different backbones. The different experiments showed the possibility to classify different materials thanks to DL models and HSI, however some difficulties in materials classification arose concerning materials that have similar properties. The preprocessing step does impact a lot on the performance of the models: applying the logarithmic derivative gave less good results compared to applying hyper-hues (HH) as a preprocessing step. During the experiments, we noticed that ResUNet had a F1-score of 88.36% . Further research can be done by detecting degradation levels of materials and detecting materials of other buildings with more data.