Meaningful Labeling of Satellite Image Details: Recent Developments in Latent Dirichlet Allocation Topics

Abstract

During the last years, we have already seen a number of attempts to use Latent Dirichlet Allocation (LDA) for local image patch classification. We can demonstrate that the reliability of these conventional labelling techniques can still be improved by more meaningful and robust label definitions derived from scene-dependent appropriate combinations of Latent Dirichlet Allocation topics for each label. This is a practical and innovative alternative to sometimes more complicated multi-labels that have been proposed by other authors. Another decided advantage of this adaptive strategy is that we can maximize the identification and discrimination capabilities of satellite image classifiers, in particular for multi-band imagery as provided by typical modern optical or radar instruments. The resulting semantic categories that we expect as classification results fulfill our initial expectations as we can exploit the full information content of the given data. We will describe how freely available and well-calibrated data of the European Sentinel-1 and Sentinel-2 satellite missions can be transformed into semantic content maps. In our case, we concentrate on remote sensing images depicting recent volcano eruptions, forest fires, floods, and ice loss in glaciers/icebergs, and on dedicated optimizations of numerical Latent Dirichlet Allocation parameters. Our results demonstrate that optimized Latent Dirichlet Allocation represents an interesting alternative to many other image classification techniques.