The creation of geographical maps using satellite images is a supervised classification problem. This means that algorithms are trained to identify patterns of interest from labeled pixels (samples). Deep learning-based techniques have achieved significant advances in this area, but several computational challenges remain, especially when the models are applied to real-world problems with class imbalance, under-representation, and unknown classes. Additionally, pixel annotation is costly and requires expertise in the target application, limiting the volume of available annotated data.
This project aims to tackle these computational challenges to increase the robustness of satellite image classification models. The developed techniques will be evaluated in real-world problems, such as mapping rural roads and monitoring urban housing conditions and their relationship with dengue outbreaks.