Improving change detection performance with generative-adversarial augmentation of dataset
Keywords: Change detection, unmanned aerial vehicle, maps updating, neural networks, generative adversarial learning
Abstract. The relevance of maps is a prerequisite for most geospatial applications such as navigation, urban planning, cadastre updating, etc. The main source of information used for maps updating is remote sensing, and the progress in sensors and methods of data analysis allows automatically retrieving the changes in observed scene from multi-time image series. Nowadays unmanned aerial vehicles (UAV) became a readily available and power mean for acquiring aerial images of a given territory. But solving change detection task using UAV-acquired images has some additional specificity, caused by additional disturbing specifics. This study addresses the problem of improving change detection performance in UAV-acquired imagery. The proposed approach firstly provide accurate image registration based on orthophoto generation, and then uses special technique for augmentation of data, that allows to improve the performance of the network model for change detection. The evaluation of the developed framework on the collected UAV-acquired multi-temporary image dataset has demonstrated change detection improving.