Super-Resolution Algorithm Applied in the Zoning of Aerial Images

Nowadays, multiple applications based on images and unmanned aerial vehicles (UAVs), such as autonomous flying, precision agriculture, and zoning for territorial planning, are possible thanks to the growing development of machine learning and the evolution of convolutional and adversarial networks. Nevertheless, this type of application implies a significant challenge because even though the images taken by a high-end drone are very accurate, it is not enough since the level of detail required for most precision agriculture and zoning applications is very high. So, it is necessary to further improve the images by implementing different techniques to recognize small details. Hence, an alternative to follow is the super-resolution method, which allows constructing an image with the information from multiple images. An efficient tool can be obtained by combining drones' advantages with different image processing techniques. This article proposes a method to improve the quality of images taken on board in a drone by increasing information obtained from multiple images that present noise, vibration-induced displacements, and illumination changes. These higher resolution images, called super-resolution images, allow supervised training processes to perform different zoning methods better. In this study, GAN-type networks show the best results to recognize visually differentiated ones on an aerial image automatically. The quality measure of the super-resolution image obtained by different methods was defined using sharpness and entropy metrics, and a semantic confusion matrix measures the accuracy of the following semantic segmentation network. Finally, the results show that the super-resolution algorithm's implementation and the automatic segmentation provide an acceptable accuracy according to the defined metrics.