Separation of Closely Located Buildings on Aerial Images Using U-Net Neural Network

Deep learning and modern type of neural network technologies are increasingly used for the detection, segmentation and classification of different objects in aerial multichannel images. The goal of given research was to develop a deep learning algorithm for automated building detection on four-channel satellite images. It is proposed to use U-Net neural network with two decoders to separate objects, one decoder is trained to segment buildings and structures, and the other detects narrow distances between buildings. It is shown that optimized U-Net can be used to detect such kind of objects efficiently. The model was implemented by means of open Keras library and launched on modern GPUs of high-performance supercomputer NVIDIA DGX-1. Before testing of developed algorithm on Planet aerial image dataset, modified U-Net had been pre-trained on SpaceNet database. The process of image data augmentation is described. The problem of effective building detection on high-resolution aerial photos can be used for urban planning, building control, search of the best locations for future outlets etc.