Towards reliable monitoring of grass species in nature conservation: Evaluation of the potential of UAV and PlanetScope multi-temporal data in the Central European tundra
Abstrakt
The reliability of remote sensing technology is essential for its efficient operational applications in nature conservation practices. In our study, multi-temporal, multispectral, and hyperspectral data from an Unoccupied Aerial Vehicle (UAV) with a ground sampling distance (GSD) of 3 cm were applied to a permanent research plot (1 ha), while spaceborne PlanetScope multi-temporal data with GSD of 3 m were used for a wider area of 250 ha.
The main aim of the study was to evaluate the potential of these data for monitoring vegetation at the species level in the Central European relict arcto-alpine tundra (in the Krkonoše Mountains, Czech Republic). We evaluated both types of data and their performance for various classifiers, and the benefits of adding multitemporal aspects to the classification and robustness of remote sensing for monitoring grass species.
This study addresses the requirements of nature conservation practices in the Krkonoše Mts. National Park to find a suitable method for monitoring changes in dominant and expanding grass species in response to global warming.
The object-based Support Vector Machine, pixel-based Random Forest and Maximum Likelihood classifiers were evaluated. We confirmed that the UAV multispectral and hyperspectral data with GSD of 3 cm can monitor dominant grass species (Nardus stricta, Calamagrostis villosa, Molinia caerulea, Deschampsia cespitosa) in tundra with excellent accuracy (95.9% and 94.0% overall accuracy for multi-temporal and mono-temporal composites, respectively).
The best F1-scores for these species ranged from 94.7 to 99.4%. However, the accuracies were substantially lower for small and sparse growing species (Avenella flexuosa and Carex bigelowii).
Unexpectedly, the UAV multispectral and hyperspectral data yielded comparable results. The multi-temporal approach improved the accuracy of all types of data.
Overall, we confirmed the robustness of the remote sensing approach for grass monitoring at the species level given the high spatial match between various classification outputs. Our study established a workflow for reliable monitoring of tundra grass vegetation according to the requirements of nature conservation and proposed innovative measures for the spatial accuracy assessment of classification outputs.
