Comparing multispectral and hyperspectral UAV data for detecting peatland vegetation patterns

Abstract

Northern peatland vegetation exhibits fine-scale spatial and spectral heterogeneity that can potentially be captured with uncrewed aerial vehicle (UAV) data due to their ultra-high spatial resolution (<10 cm). From this perspective, the contribution of different spectral sensors in mapping various vegetation characteristics has not been thoroughly investigated. We delineated spatial patterns of plant community clusters, plant functional types (PFTs), and selected plant species with UAV hyperspectral (HS), UAV multispectral (MS), and airborne LiDAR (light detection and ranging) topography (TP) data in two northern peatlands. We conducted random forest (RF) regressions in a geographic object-based image analysis (GEOBIA) framework and compared the relative contributions of the different datasets. In the best regression models, the percentage of explained variance was 24–74 % (RMSE:0.24–0.31), 40–90 % (RMSE:0.12–0.41), and 18–90 % (RMSE:0.03–0.40) for plant community clusters, PFTs, and plant species, respectively. The MS-TP combination had, in many cases, the highest performance, while HS-based models had better performance for some plant community clusters, PFTs, and plant species. TP features were useful only in certain situations. Overall, our results suggest that UAV MS imagery combined with TP data outperformed or performed at least almost as well as the models using UAV HS data and while all data combinations are capable for fine-scale detection of vegetation in northern peatlands. A more comprehensive investigations of data processing and methodology selection is needed to conclude if there is an added value of UAV HS data for peatland vegetation monitoring.

Keywords

• Peatland vegetation mapping
• Hyperspectral remote sensing
• Geographic object-based image analysis
• Random forest