Terrain change monitoring from multi-temporal aerospace images using advanced training neural networks

One of the key elements of the digital map making process is formation and updating of objects by detecting terrain changes from multi-temporal highly detailed aerospace images. At the same time, at present the main predominant method of change detecting the said images is still their direct visual analysis, implemented by the deciphering operator of the ground complex of view information processing, which, in general, reduces the efficiency of digital maps preparation. Application of progressive technologies of automatic change detection will allow ensuring a better quality of digital maps, including those according to heterogeneous remote sensing data (radar imagery, laser scanning). Methods based on artificial intelligence technologies, in particular advanced training of neural networks, show significant ‘stability’, which is due to formation of the most informative signs of significant changes in the process of neural network training. The proposed developed ResNet-50-CDB neural network and algorithm for reversible processing of input images will provide lower probability of errors of the first and second type, which gives both higher Jaccard index and accuracy of boundary localisation