Semantic segmentation of earth scanning images using neural network algorithms

The relevance of the research is caused by the need to develop methods, algorithms and software to improve the effectiveness of semantic segmentation of the Earth scanning images. Today there is a need to improve the quality of semantic segmentation of objects in images, despite the intensive development of modern methods and algorithms, often they do not provide the required quality of work and reliability. The main aim of the research is to develop the algorithms to solve the problem of semantic segmentation of the Earth scanning images. Objects: neural network algorithms which provide semantic segmentation of images; methods of implementation and training of artificial neural networks; image processing algorithms. Methods. To solve the tasks, the authors have used the methods of computational Intelligence; methods of pattern classification on images; theory of artificial neural networks; methods for training of artificial neural networks using Visual Studio software; deep learning framework Caffe for implementation of neural network algorithms. Results. The authors made a review of methods and algorithms which allow carrying out semantic segmentation of images. Based on the analysis, it was concluded that neural network algorithms provide more efficient results. The authors developed the convolutional neural network with the original architecture consisting of six layers. Software implementation of the described algorithms is implemented. It allows building a map of segmented buildings, roads and background based on input data. The paper introduces the comparison of results of using different learning algorithms for the developed neural network.