ATTRIBUTE PROFILES WITHOUT THRESHOLDS THROUGH HISTOGRAM BASED TREE PATH DESCRIPTION

Despite the repeatedly validated superiority of deep learning methods compared to shallow pixel desription techniques of the near past, they are heavily dependent on the abundance of training data; of which unfortunately the remote sensing field never possesses enough. Consequently, methods such as morphological attribute profiles are still among the best available options for spectral-spatial description. In this paper, we focus on one of their few yet significant drawbacks: the need for a predefined threshold set. We propose to describe every pixel through the histogram of attribute values in the tree path containing it. We validate our method with two hyperspectral datasets and two attributes, and show that it exhibits either comparable or superior performance with respect to manual or automatically calculated thresholds.