Hyperspectral vs. Multispectral data: Comparison of the spectral differentiation capabilities of Natura 2000 non-forest habitats

Identification of the Natura 2000 habitats using remote sensing techniques is one of the most important chal-lenges of nature conservation. In this study, the potential for differentiating non-forest Natura 2000 habitats from the other habitats was examined using hyperspectral data in the scope of VNIR (0.4-1 mu m), SWIR (1-2.5 mu m) and simulated multispectral data (Sentinel-2). The aim of the research was also to determine the most informative spectral ranges from the optical range. Five different Natura 2000 habitats common in Central Europe were analysed: heaths (code 4030), mires (code 7140), grasslands (code 6230) and meadows (codes 6410 and 6510). In order to guarantee the objectivity and transferability of the results each habitat was tested in two areas and in three campaigns (spring, summer, autumn). Hyperspectral data was acquired using HySpex VNIR-1800 and SWIR-384 scanners. The Sentinel-2 data was resampled based on HySpex spectral reflectance. The overflights were performed simultaneously with ground reference data - habitats and background polygons. The Linear Discriminant Analysis was performed in iterative mode based on spectral reflectance acquired from hyperspectral and multispectral data. This resulted in distribution of correctness rate values and information about the most differentiating spectral bands for each habitat. Based on the results of our experiments we conclude that: (i) hyperspectral data (both VNIR and SWIR) obtained from May to September was useful for differentiation of habitats from background with efficiency reaching over 90%, regardless of the area; (ii) the most useful spectral ranges are: in VNIR -0.416-0.442 mu m and 0.502-0.522 mu m, in SWIR -1.117-1.165 mu m and 1.290-1.361 mu m; (iii) the potential of multispectral data (Sentinel-2) in distinguishing Natura 2000 habitats from the background is diverse; higher for heaths and mires (comparable to hyperspectral data) lower for meadows (6410, 6510) and grasslands (6230); (iv) in case of meadows and grasslands, the correctness rate for the Sentinel-2 data was on average about 20% lower compared to the hyperspectral data.

Automated summary

Identification of Natura 2000 habitats using remote sensing techniques is a significant challenge in nature conservation. This study examined the potential for differentiating non-forest Natura 2000 habitats from other habitats using hyperspectral and multispectral data. The research identified the most informative spectral ranges and concluded that hyperspectral data from May to September was useful for differentiation with efficiency over 90%. Multispectral data showed varied potential in distinguishing habitats, with heaths and mires performing better than meadows and grasslands.