A Real-Time Aerial Semantic Segmentation System Based on U-Net Deep Learning Using Drone Images

Image segmentation at the pixel level is a time-consuming and difficult task in computer vision and image processing. Aerial (satellite/drone) photo segmentation is considered in this paper. Data from high-resolution remote sensing has enabled new applications such as more detailed per-pixel object classification. U-Net with VGG16 has made segmentation and categorization of images much more efficient and intelligent. U-Net models with pre-trained VGG16 backbones perform best across all tested scenarios. Adding the near-infrared band improves prediction results slightly compared with using RGB image bands alone. The ability to transfer images between sensors, especially between satellites and aerial images, could be improved through train-time enhancement and contrast enhancement. Further improving performance could be achieved by adding noisy training data from free online resources.