Development Of An Intelligent Analysis, Search, And Recognition Method For Anthropogenic Target Objects In Remote Sensing Data Using Generative Adversarial Networks

Abstract

This paper presents an algorithm for the semantic segmentation of image regions containing anthropogenic objects. The method is based on the use of generative adversarial networks (GANs). The proposed detector operates by identifying a specific spectral artifact that arises when image resolution is enhanced using a generative adversarial network.

The study analyzes a spectral neural network-based detector that relies on detecting characteristic spectral artifacts produced by the resolution enhancement module of generative neural networks. This detector was chosen for investigation because its developers claim high efficiency in analyzing various types of images, including the detection of artificially generated images obtained via remote sensing (RS). However, previous research has only tested this detector on fully synthesized images. It remains unclear how effective it would be in scenarios where only specific regions of RS images require refinement. Additionally, prior studies utilized only a single type of satellite imagery, and the generator set did not account for adversarial networks with a fixed long-element generator.

This research focuses on studying semantic analysis algorithms for Earth observation satellite imagery. Currently, numerous image segmentation methods exist, and the development of algorithms for processing remote sensing data continues. Some of these methods are based on generative adversarial neural networks.

The proposed algorithm is designed for the segmentation of remote sensing images of varying types and resolutions. The processed image fragments differ in size, shape, and may be non-contiguous.