Deep learning in very high resolution remote sensing image information mining communication concept

This paper presents the image information mining based on a communication channel concept. The feature extraction algorithms encode the image, while an analysis of topic discovery will decode and send its content to the user in the shape of a semantic map. We consider this approach for a real meaning based semantic annotation of very high resolution remote sensing images. The scene content is described using a multi-level hierarchical information representation. Feature hierarchies are discovered considering that higher levels are formed by combining features from lower level. Such a level to level mapping defines our methodology as a deep learning process. The whole analysis can be divided in two major learning steps. The first one regards the Bayesian inference to extract objects and assign basic semantic to the image. The second step models the spatial interactions between the scene objects based on Latent Dirichlet Allocation, performing a high level semantic annotation. We used a WorldView2 image to exemplify the processing results.