WildPixels: dense labeling of remote sensing images “in the wild” is a bold, interdisciplinary research project to make large-scale geographic mappings through supervised learning from a few annotated pixels. The study aims to develop new computational approaches that fill some gaps that exist between recent literature in the area and what real-world applications really require. Among the applications, three stand out: detection of rural roads in the Amazon and cerrado; mapping of urban areas with potential for Dengue infestation, in addition to the recognition of native species and indicators of climate change.

The main challenge for effectively using supervised learning for dense pixel labeling is the lack of robustness of the models when applied in nature. There are already great advances in deep learning, but they need a lot of annotated data and, to learn patterns for application in the real world, there is a need for annotations from experts or rare phenomena joint to the research, which can bring problems of imbalance, scarcity and/or or noise in the data. This is something extremely complex to solve from the point of view of machine learning, that is, from the point of view of Computer Science, which these researches overcome and advance.

Through studying and developing new approaches to increase the robustness of models to these constraints, researchers have focused on critical machine learning problems at the pixel level: (1) class imbalance; (2) underrepresented classes (few-shot learning); and (3) identification of classes and objects not seen in the training data (open set recognition), providing robust and validated results.