Automatic segmentation of satellite imagery based on the modified UNET convolutional neural network
Abstract
Automatic segmentation of satellite imagery based on the modified UNET convolutional neural network
Incoming article date: 25.10.2017The article suggests a technique for automatic segmentation of satellite images on the basis of convolutional neural networks into several classes, such as buildings, rivers, roads, etc. The software implementation of the proposed methodology took the second place in the competition for the segmentation of satellite imagery on the Kaggle platoform in competition: Dstl Satellite Imagery Feature Detection. The article describes how to prepare images for the training of neural network and reveal details for full dataflow and the principles of the traning. The structure of the neural network for segmentation is proposed. The network is built on the basis of UNET with additional BatchNormalization and Dropout layers, based on double convolution blocks. A procedure for cross-evaluation is described to assess the accuracy of the models obtained. The descriptions of algorithms for postprocessing and the technique of segmentation refinement are presented by using an ensemble of several models. A specialized model is proposed for finding objects of small size, such as "cars" and "motorcycles". An overview of other methods used to solve this problem is also given, which were not included in the final solution. In the experimental results it is shown that the efficiency of neural networks in this task is extremely high and it is possible to automatically prepare a layout of the terrain similar to the markup made by human. And thereby it allows to save money, since significant financial resources are being spent on manual marking.
Keywords: convolutional neural nets, sattelite imagery, image segmentation, machine learning, crossvalidation, Jaccard coefficient, UNET network, image classification, computer vision, contest results