Selection of spectral bands for the detection of dedicated semantic classes in satellite disaster images by Latent Dirichlet Allocation

Kurzfassung

When we want to detect and classify fires in multispectral satellite images (e.g., various types of forest fires and their consequences), we are faced with the prob-lem of how to recognise and interpret the visible fire patterns, and how to dis-criminate fires from smoke and clouds (and other less relevant classes). Current-ly preferred image classification techniques often rely on deep learning (DL) ap-proaches calling for extensive training efforts with lots of training data. Thus, we must apply other techniques if no such data are readily available. In this latter case, we can resort to Latent Dirichlet Allocation (LDA), a dictionary-generated image topic-based classification approach, where no explicit training and learn-ing steps need to be added. Then, the correct assignment of semantic meaning to individual image patches can be accomplished by a priori image knowledge and some off-line image interpretation provided by expert users. In the following, we describe a Latent Dirichlet fire detection and visualisation approach and its per-formance based on an optimised selection of Sentinel-2 satellite spectral image bands. The combination of automated image classification and multi-colour vis-ualisation seems to be an interesting alternative to deep learning. Thus, our paper supports “data to knowledge”.