Aerial pictures semantic segmentation applying deep learning

An obvious expansion in the measure of satellite dataset accessible lately has made the translation of this information a difficult issue at scale. Determining helpful insights from such pictures requires a rich comprehension of the data present in them. AI is currently utilized for keeping up precise automated regional maps to react to real time, natural and catastrophe recuperation challenges. These assignments need close to continuous, precise, mechanized planning straight from aerial and satellite pictures. In this project, we apply Mask-RCNN and Conditional Adversarial Network techniques for extracting building footprint. The problem is viewed as a supervised learning problem. We try different things with learning parameters and algorithms, apply data augmentation, use transfer learning, utilizing RGB data and to accomplish high precision results. The resulting pipeline incorporates image pre-processing algorithms that permits it to adapt to input pictures of fluctuating quality, resolution and channels.