Possibilities of convolutional neural networks use for remote sensing image classification
Ondřej Pešek · 2024
In recent years, the technological progress in certain science fields is getting faster and faster. The advancement acceleration makes it challenging for other scientific areas to keep up with this tempo. One of the~exemplary relationships is the link between convolutional neural network development and the province of geomatics or remote sensing. New architectures of convolutional neural network models are being published with an expeditious tempo, not leaving remote sensing scientists enough time to run thorough analyses and comparisons of their performance. As a result, many remote sensing studies tend to use the most recent architectures, although the knowledge of the architectures' relative performance is limited to their original scientific field. However, models useful for common computer vision problems do not necessarily have to reach good results on remote sensing data. The aim of this thesis is to perform systematic research on the possibilities of use of chosen convolutional neural network architectures on three selected use cases from the field of remote sensing---cloud detection on VENuS satellite system imagery, land use classification of Prague urban green areas on Sentinel-2 satellite system imagery, and road surface classification on aerial imagery.