Statistical approaches based on Extreme Value Theory (EVT) and Variational Inference (VI) can be effective in dealing with supervised learning in the context of dense labeling of remote sensing images. These techniques have already been used effectively in OSR or few-shot learning in sparse classification tasks, but have not yet been explored for semantic segmentation. Existing algorithm models, despite having some effectiveness, are not yet prepared to analyze large-scale data, using high-resolution images over large areas, especially in places with little annotated data. Thus, the main objective of this project is to develop new methods based on EVT and VI for dense tagging of remote sensing images with sparse annotation. In this way, evaluate the methods developed in practical and more complex situations.
At the same time, we are developing new learning approaches for little annotated training data; new open set semantic segmentation methods; evaluating different training methodologies for different types of user annotation, including weakly labeled pixels; exploring new forms of interactive learning; and developing new approaches taking into account aspects of geographic scalability and domain change.
We hope to contribute to the production and promotion of knowledge through quality publications; develop new scientific methodologies both in theory and in practice; train and qualify students; in addition to promoting new technologies with the potential to bring economic, social and environmental benefits.
