Domain Aware Learning for Hyperspectral Image Classification using Spectral-Spatial Features from Externally Labelled Data

One of the major challenges in the classification of remotely sensed hyperspectral data is the limited number of labelled pixels within an image. This means that the labelled and unlabelled pixels can have different distributions, leading to poor performance by semi-supervised machine learning models. Our work addresses this problem of domain shift, and shows how a domain aware model can be used to overcome this problem. We employ our model on the Pavia University and Pavia Centre dataset, and compare the performance of our model with a traditional deep neural network model. We observed an improvement of up to 10 percent in the classification accuracy. We also analysed the effect of spatial and spectral features on our domain aware model, and came to a basic conclusion that spectral features provide a more global information about a pixel’s class, whereas spatial features provide a more local information, which means that domain shifts are likely to be more pronounced in spatial data as compared to spectral data.