Deep Learning for Satellite Image Analysis

Deep learning architectures have the potential of saving the world from losing football fieldsized forest areas each second. These architectures possess large learning capacities when compared to conventional machine learning architectures, and thus are trained on sizable data-sets to efficiently extract both coarse and fine features from various image scenes. As a result, they can provide crucial information that is needed to manage the deforestation process and its consequences on the environment and ecosystem more effectively. This thesis outlines the two deep learning based systems designed for satellite image analysis. The first system analyzed satellite images of the Amazon, and the goal was to interpret the image content by providing a set of labels that best describe it. The highest performing architecture was able to achieve a score of 92.886% while a combination of several high performance, yet uncorrelated, architectures increased the overall score to 93.070%. This result is only 0.248% lower than what current state of the art algorithms achieved on the same task. The second system was designed to detect the presence of clouds in Landsat 8 images by analyzing small chips within each large image. This system produced cloud masks, which were then compared to the corresponding ground truth cloud masks obtained from the provided images. The predicted cloud masks were able to achieve an average score of 92.931%, which is very high for the given accuracy measure.