Airborne hyperspectral imaging has great potential for the land cover mapping of large urban areas. As much as artificial impermeable surfaces impact watershed hydrology (particularly droughts and floods), urban heat island effects and soil carbon uptake, providing public authorities with detailed land cover maps is a key issue to mitigate the effects of urban sprawl. However, the use of state-of-the-art machine learning algorithms to automatically map the land cover has been dramatically limited by the scarcity of training data (prohibitively expensive to acquire through field campaigns) with respect to the diversity of ground materials and to their large spectral intra-class variability. Therefore, we investigate in this work the introduction of deductive biases derived from a priori physical knowledge into representation learning in order to improve the generalizability of classification models in a weakly supervised regime. Precisely, we integrate into a latent generative model physical layers that deterministically model some of the true underlying factors of variation of the data. We discuss the benefits of our method through numerical experiments on simulated and real data sets in terms of classification accuracy, interpretability and disentanglement.
