| Abstract [eng] |
Remote sensing techniques can greatly improve the efficiency of field surveys by enabling to remotely collect information quickly and conveniently. Using satellite imagery, unmanned aerial vehicles and other remote sensing techniques, researchers can facilitate data collection and classification. The remote-sensing system being developed in this thesis is designed to facilitate the monitoring of forest cover and the classification of river boulders using remote-sensing techniques. The development of the system leads to three studies: the suitability of automatically generated datasets for training forest classification models, the impact of different combinations of satellite spectral bands on forest classification models, and a study of the feasibility of classifying river boulders. In the first study, it was found that automatically generated datasets from openly available data allow to achieve high accuracy classification results (~0.92 PA and ~0.84 mIoU). The second study found that different combinations of spectral bands can influence classification patterns. The combination of near-infrared and visible light achieved the highest accuracy (0.96 PA and 0.88 mIoU). Finally, a study on the classification of river boulders indicated high classification accuracy (~0.776 P) and average recall rates (~0.690 R) for river boulders over water. On the other hand, the classification of boulders under water showed an average precision results (~0.633 P) but low recall rates (~0.391 R). |
