APCRS-2021-0002
AUTHOR(S):
Kamil Bartosik, Zdzisław Kurczyński
TITLE:
USING CONVOLUTIONAL NEURAL NETWORKS TO IMPROVE THE SPATIAL RESOLUTION OF AERIAL IMAGES
ABSTRACT:
Artificial intelligence is rapidly finding increasing popularity and applications. In the field of photogrammetry and remote sensing, convolutional neural networks (CNNs) are applied, especially for object detection and classification in aerial and satellite images. Some variations of such networks are super-resolution convolutional neural networks (SRCNNs). Proposed only a few years ago, they are already finding applications for increasing the spatial resolution of aerial and satellite images. This problem is addressed in the paper. The authors introduce the issue of convolutional neural networks and, on this background, the specifics of super-resolution convolutional neural networks. In the experimental part, they undertook the task of improving the resolution of high-resolution aerial images acquired with a large-format camera for a city area with original resolution (ground sampling distance) GSD=5 cm. These images were degraded to varying degrees and then subjected to resolution enhancement with the help of a constructed network. The resulting (sharpened) images were evaluated quantitatively using defined measures of resolution improvement and visual comparison with the original images. The results, surprisingly good especially for images with twice the resolution degradation, confirm the great potential of convolutional neural networks to improve the spatial resolution of aerial images is confirmed.
KEY WORDS: convolutional neural networks, high-resolution convolutional neural networks, image spatial resolution, image resolution enhancement
DOI:http://doi.org/10.14681/apcrs-2021-002