The value of hyperspectral UAV imagery in characterizing tundra vegetation
Putkiranta, Pauli; Räsänen, Aleksi; Korpelainen, Pasi; Erlandsson, Rasmus; Kolari, Tiina M. H.; Pang, Yuwen; Villoslada, Miguel; Wolff, Franziska; Kumpula, Timo; Virtanen, Tarmo (2024)
Putkiranta, Pauli
Räsänen, Aleksi
Korpelainen, Pasi
Erlandsson, Rasmus
Kolari, Tiina M. H.
Pang, Yuwen
Villoslada, Miguel
Wolff, Franziska
Kumpula, Timo
Virtanen, Tarmo
Julkaisusarja
Remote sensing of environment
Volyymi
308
Elsevier
2024
Cite: Pauli Putkiranta, Aleksi Räsänen, Pasi Korpelainen, Rasmus Erlandsson, Tiina H.M. Kolari, Yuwen Pang, Miguel Villoslada, Franziska Wolff, Timo Kumpula, Tarmo Virtanen, The value of hyperspectral UAV imagery in characterizing tundra vegetation, Remote Sensing of Environment, Volume 308,2024,114175,ISSN 0034-4257, https://doi.org/10.1016/j.rse.2024.114175.
Julkaisun pysyvä osoite on
http://urn.fi/URN:NBN:fi-fe2024051530827
http://urn.fi/URN:NBN:fi-fe2024051530827
Tiivistelmä
The fine-scale spatial heterogeneity of low-growth Arctic tundra landscapes necessitates the use of high-spatial-resolution remote sensing data for accurate detection of vegetation patterns. While multispectral satellite and aerial imaging, including the use of uncrewed aerial vehicles (UAVs), are common approaches, hyperspectral UAV imaging has not been thoroughly explored in these ecosystems. Here, we assess the added value of hyperspectral UAV imaging relative to multispectral UAV imaging in modelling plant communities in low-growth oroarctic tundra heaths in Saariselkä, northern Finland. We compare three different spectral compositions: 4-channel broadband aerial images, 5-channel broadband UAV images and 112-channel narrowband UAV images. Based on field vegetation plot data, we estimate vascular plant aboveground biomass, leaf area index, species richness, Shannon's diversity index, and community composition. We use spectral and topographic information to compile 12 explanatory datasets for random forest regression and classification.
For aboveground biomass and leaf area index, the highest R2 values were 0.60 and 0.65, respectively, and broadband variables were most important. In the best models for biodiversity metrics species richness and Shannon's index R2 values were 0.53 and 0.46, respectively, with hyperspectral, topographic, and multispectral variables having high importance. For 4 floristically determined community clusters, both random forest classifications and fuzzy cluster membership regressions were conducted. Overall accuracy (OA) for classification was 0.67 at best, while cluster membership was estimated with an R2 of 0.29–0.53. Variable importance was heavily dependent on community composition, but topographic, multispectral, and hyperspectral data were all selected for these community composition models. Hyperspectral models generally outperformed multispectral ones when topographic data were excluded. With topographic data, this difference was diminished, and performance improvements from added hyperspectral data were limited to 0–10 percentage point increases in R2, the largest occurring in the metrics with lowest R2. These results suggest that while hyperspectral can outperform multispectral imaging, multispectral and topographic data are mostly sufficient in practical applications in tundra heaths.
For aboveground biomass and leaf area index, the highest R2 values were 0.60 and 0.65, respectively, and broadband variables were most important. In the best models for biodiversity metrics species richness and Shannon's index R2 values were 0.53 and 0.46, respectively, with hyperspectral, topographic, and multispectral variables having high importance. For 4 floristically determined community clusters, both random forest classifications and fuzzy cluster membership regressions were conducted. Overall accuracy (OA) for classification was 0.67 at best, while cluster membership was estimated with an R2 of 0.29–0.53. Variable importance was heavily dependent on community composition, but topographic, multispectral, and hyperspectral data were all selected for these community composition models. Hyperspectral models generally outperformed multispectral ones when topographic data were excluded. With topographic data, this difference was diminished, and performance improvements from added hyperspectral data were limited to 0–10 percentage point increases in R2, the largest occurring in the metrics with lowest R2. These results suggest that while hyperspectral can outperform multispectral imaging, multispectral and topographic data are mostly sufficient in practical applications in tundra heaths.
Collections
- Julkaisut [85849]