Аннотация:In this paper, the problem of neural-based detection of vehicles on blurred and noisy images of remote sensing is considered. The aim of this study is to improve the performance of object detection in aerial images exposed to distortions of various nature. This aim is achieved using the proposed technique, which includes optimal training images augmentation and an adaptive input data preprocessing reducing the impact of high-frequency noise. The simulation results, carried out on the publicly available DOTA aerial images dataset, show that the application of optimal robust training and the use of the proposed adaptive system simultaneously allow to obtain a more robust neural-based vehicle detection system for various tasks of remote sensing.