Assessing floristic composition with multispectral ... - Semantic Scholar
International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
Contents lists available at SciVerse ScienceDirect
International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag
Assessing floristic composition with multispectral sensors—A comparison based on monotemporal and multiseasonal field spectra Hannes Feilhauer a,b,c,∗ , Frank Thonfeld b , Ulrike Faude c,f , Kate S. He d , Duccio Rocchini e , Sebastian Schmidtlein b,c a
Department of Geography, University of Erlangen-Nuremberg, Kochstr. 4/4, D-91054 Erlangen, Germany Center for Remote Sensing of Land Surfaces, University of Bonn, Walter-Flex-Str. 3, D-53113 Bonn, Germany c Vegetation Geography, University of Bonn, Meckenheimer Allee 166, D-53115 Bonn, Germany d Department of Biological Sciences, Murray State University, 2112 16th Street, Murray, KY 42071, USA e Fondazione Edmund Mach, Research and Innovation Centre, Department of Biodiversity and Molecular Ecology, GIS and Remote Sensing Unit, Via E. Mach 1, 38010 S. Michele all’Adige (TN), Italy f EFTAS Fernerkundung, Oststraße 2-18, 48145 Münster, Germany b
a r t i c l e
i n f o
Article history: Received 30 July 2012 Accepted 10 September 2012 Keywords: Conservation Isometric Feature Mapping Optical remote sensing Partial Least Squares regression Subclass information Vegetation
a b s t r a c t Assessing and mapping patterns of (semi-)natural vegetation types at a large spatial scale is a difficult task. The challenge increases if the floristic variation within vegetation types (i.e., subtype variation of species composition) is the target. A desirable way to deal with this task may be to address such vegetation patterns with remote-sensing approaches. In particular data from multispectral sensors are easy to obtain, globally accessible, and often provide a high temporal resolution. They hence offer a comprehensive basis for vegetation mapping. The potential of such sensors for vegetation mapping has, however, never been thoroughly investigated. In particular, a systematic test regarding the spectral capabilities of these data for an assessment of detailed floristic variation has not been implemented to date. We thus addressed in this study the question how the ability of optical sensors to map floristic variation is affected by their respective spectral coverage and number of bands. To answer this question, we simulated monotemporal and multiseasonal data of eleven multispectral sensors. These data were used to model gradual transitions in species composition (i.e., floristic gradients) within three types of spontaneous vegetation typical for Central Europe using Partial Least Squares regression. Comparison of the model fits (ranging up to R2 = 0.76 in cross-validation) illustrated the potential of multispectral data for detailed vegetation mapping. The results show that spectral coverage of the entire solar-reflective domain is the most important sensor characteristic for a successful assessment of floristic variation. Model and sensor performances as well as limitations are thoroughly discussed, and recommendations for sensor development are made based on the final conclusions of this study. © 2012 Elsevier B.V. All rights reserved.
1. Introduction Maps of land-cover and vegetation types are crucial for decisionmaking in conservation management and planning (Leemans and Zuidema, 1995). These maps are frequently based on optical remote-sensing data acquired by air- or spaceborne sensors (Xie et al., 2008). Since satellite imagery offers the opportunity to depict the state of large areas rapidly, time series of derived mapping products may further represent a viable basis to monitor a dynamically changing landscape on a regular basis (e.g., Coops et al., 2009). Users interested in such monitoring opportunities frequently require detailed subtype information on habitat conditions (i.e., the
∗ Corresponding author. Tel.: +49 9131 85 26680. E-mail address: [email protected] (H. Feilhauer). 0303-2434/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jag.2012.09.002
fine-scale variability of a type). The requested information may include the presence of invasive species or degradation indicators, structural attributes such as shrub and grass encroachment, or the species composition of a vegetation type. In terms of rather coarse land-cover classes, these variations will be undetected due to the generalization. For conservation planning, this information may, however, be decisive for the assessment of a site in order to meet the standards of, e.g., the European Union directives on the conservation of natural habitats and of wild fauna and flora, a program aiming to protect threatened habitats and species across Europe (Council of the European Communities, 1992). In consequence, various practical considerations have been made on how remote sensing can meet this demand (see, e.g., Vanden Borre et al., 2011). Multispectral data taken by spaceborne sensors are easy to obtain for land-cover assessments of large areas. Such data have also been used to assess habitat conditions for monitoring
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
purposes (Förster et al., 2008; Franke et al., 2012; Spanhove et al., 2012; Vogelmann et al., 2012) and to detect invasive species (e.g., Dorigo et al., 2012; Ghioca-Robrecht et al., 2008); however, these assessments seldom include floristic variation within discrete habitats. Although some studies investigate spectral characteristics of specific sensors for such tasks (e.g., Schuster et al., 2012), a systematic test of the suitability of multispectral data for an assessment of detailed floristic variation has not been implemented up to now. To address this gap, we tested in this study to what extent the spectral information of multispectral data allows for a detailed assessment of the variable floristic composition within a certain natural habitat. We further tested which sensor characteristics are responsible for success or failure in this task. 2. Materials and methods 2.1. Study sites and sampling design To answer these questions, an empirical study was implemented in three different types of spontaneous vegetation very common in Central Europe: a nutrient-poor and dry grassland, a wet heath in transition to mire, and a nutrient rich and wet floodplain meadow. For the aim of this study, we tested the potential of eleven multispectral sensors to spectrally discriminate these vegetation types and to assess the within-type floristic variation with simulated data in a multivariate approach (see Fig. 1 for a flowchart showing a scheme of the study approach). We described the variability within each vegetation type as floristic gradients (i.e., gradual transitions in species composition, Gleason, 1926) since this concept is suitable to address the fuzzy characteristics of natural vegetation.
Fig. 1. Flowchart of the analysis showing logical steps of the study approach. The vegetation records sampled in each type were subjected to an ordination to quantify and describe the variability within the type as gradual transition (i.e., floristic gradient). The signals of eleven multispectral sensors were simulated based on hyperspectral data taken in the field. Monotemporal and multiseasonal simulated data were subsequently regressed against the first ordination axes. The model fits were used to evaluate the performance of the respective sensor regarding an estimation of within-type floristic variation.
219
The three sites are located near Cologne, Germany (Fig. 2a). The nutrient-poor grassland (7◦ 07 00 E, 50◦ 54 17 N, Fig. 2b) covers an area of 26 ha. It features a floristic gradient from Agrostis capillaris–Holcus mollis grassland to stands dominated by Calamagrostis epigejos. The latter result from human disturbance. The wet heath (12 ha, 7◦ 09 57 E, 50◦ 52 27 N, Fig. 2c) includes patches of heather (dominated by the dwarf shrubs Calluna vulgaris and Erica tetralix) that are interspersed along a wetness gradient with tussocks of Molinia caerulea, stands of Narthecium ossifragum, and wet hollows with Sphagnum mosses. The floodplain meadow (12 ha, 7◦ 07 27 E, 50◦ 46 34 N, Fig. 2d) is a nutrient-rich lowland meadow with gradual transitions to tall forb communities. No land use or management activities took place during the time of data acquisition. We established permanent plots for the time of data acquisition based on a random sampling design in these three sites. The plot arrangement was constrained by minimum inter-plot distances of 5 m. Further, locations in the shade of trees or shrubs were avoided. The plots were circular in shape and covered an area of 1 m2 each. Although this plot size did not meet the requirements for an image-based analysis, it was optimized for the demands of detailed vegetation sampling and the application of field spectroscopy. Each plot center was marked with a stake during the sampling period. The nutrient-poor grassland was sampled with n = 57 plots, the wet heath with n = 35 plots, and the floodplain meadow with n = 37 plots. These three sites were subsequently treated as three vegetation types (i.e., individual classes; type and class are interchangeable in this paper). 2.2. Vegetation data per type Cover fractions of all occurring vascular plants and mosses within each plot were recorded to quantify the within type floristic variation. Cover fractions were estimated close to the vegetation optimum at the respective site. The nutrient-poor grassland was sampled on July 15th, the wet heath on July 31st, and the floodplain meadow on June 23rd, 2008. The visual estimations of cover fractions were aided by a circular 1 m2 sampling frame. Variation of species composition over the vegetation period was negligible for the study sites (Feilhauer and Schmidtlein, 2011). We hence implemented all analyses using the monotemporal vegetation records. To describe the floristic variation within each vegetation type, the vegetation data of each site were separately subjected to an ordination analysis. Ordination techniques are frequently used in ecology to describe and quantify structures in vegetation through dimensionality reduction (Legendre and Legendre, 1998). Therefore, the vegetation records are arranged numerically regarding their floristic similarity in a synthetic Euclidean space. Several algorithms are available for this task. In the present study we used Isometric Feature Mapping (Isomap, Tenenbaum et al., 2000), an ordination approach with high flexibility towards data structures. Given this flexibility, Isomap has shown promising results in previous studies (e.g., Feilhauer et al., 2011; Mahecha et al., 2007). Isomap is an ordination approach that is based on a resemblance matrix of inter-plot dissimilarities in a multidimensional floristic similarity space (Fig. 3a and b). The distance metric to quantify these dissimilarities can be chosen freely. Here we used Bray–Curtis distances, a measure well established in ecological studies (Legendre and Anderson, 1999). The Isomap algorithm is based on these distances to identify k nearest neighbors for each plot. These neighborhood relations are subsequently used to generate a new resemblance matrix with geodesic distances that connect remote points via a net of mutual neighbors (Fig. 3c).
220
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
Fig. 2. Location of the three sites in Germany (a) and plot distribution across the sites (b–d). Coordinates are given in UTM 32U WGS 84. Aerial photographs in the background depict the surface in the red spectral region with a spatial resolution of 20 cm (© underlying aerial photography: state of North Rhine-Westphalia).
The geodesic distances resulting from a small set of neighbors are highly non-linear, with a growing number of considered neighbors the linearity increases. Further, the geodesic dissimilarities are projected to lower dimensionality with a multidimensional scaling approach (Fig. 3d). Prominent non-linear structures in the data can be preserved throughout the analysis by choosing an appropriate k. We used this approach to identify the main floristic gradient (i.e., the first dimension of the Isomap space) for each site. The scores of the plots on these gradients were subsequently taken as continuous description of within-type variation of the vegetation types. Indicator species for each vegetation type were identified in a numerical analysis using the isopam algorithm (Schmidtlein et al., 2010) to illustrate this floristic variation. Isomap and isopam analyses were carried out in the R statistical environment (R Development Core Team, 2011) using the packages vegan (Oksanen et al., 2011) and isopam (Schmidtlein, 2010). 2.3. Field spectroscopy The vegetation canopy of the three sites changed during the measurement period with respect to litter fractions (i.e., dead plant material, Fig. 4b), canopy height (Fig. 4c), as well as phenological stage and drought stress. Canopy reflectance of the plots was thus measured on multiple dates (n = 7 at the nutrientpoor grassland, n = 5 at the wet heath, and n = 6 at the floodplain meadow, Fig. 4a) in the 2008 vegetation period with a fullrange field spectrometer (ASD Fieldspec 3 JRTM , ASD Inc., Boulder, CO). All measurements were taken within 2 h around solar noon relative to a SpectralonTM panel (Labsphere Inc., North Sutton, NH). These reflectance spectra covered the spectral range of 350–2500 nm in 2151 bands. The spectrometer probe was
leveled in nadir approximately 110 cm above the canopy. With the given field-of-view of 25◦ , five measurements were required to cover the whole plot area. Each measurement was taken in five repetitions. Thus, for every plot a total of 5 × 5 spectra was taken per date. We compared eleven sensors (described in Section 2.4 and Table 1) with monotemporal (i.e., one single date) and multiseasonal (i.e., multiple dates of different seasons but within a single year) data to test their performance. In practice, monotemporal analyses offer the advantage of a reduced effort and lower costs for data acquisition and preprocessing. Optimal timing of the data acquisition is crucial because the strength of statistical relations between vegetation patterns and reflectance is temporally variable. Many studies thus choose a distinct phenological stage for data acquisition (e.g., Andrew and Ustin, 2008; Ge et al., 2006; Laba et al., 2005). In an earlier study (Feilhauer and Schmidtlein, 2011), it was shown for hyperspectral data that the vegetation patterns within the sites under investigation could be most accurately differentiated during the vegetation optimum, which was defined as the date of maximum cover of photosynthetically active vegetation (i.e., the growth peak). We hence used this date for the monotemporal comparison (Fig. 4a). Yet, the configuration of various multispectral sensors (e.g., RapidEye, Sentinel-2, Worldview-2) is optimized towards a short revisit time. These sensors are predisposed to provide multiseasonal data covering phenological events continuously rather than providing a single snapshot in time. We thus quantified the additional explanatory power gained by the use of multiseasonal data considering all available time steps (i.e., the ASD sampling dates in 2008). Monotemporal and multiseasonal results gained by different sensors were compared to assess the explanatory power regarding within-type variation of vegetation.
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
221
Fig. 3. Schematic illustration of the Isomap approach. From the distribution of sample points in an n-dimensional similarity space (a), mutual inter-point distances are calculated (b). This set is used to identify the k nearest neighbors of each point (c, shown for k = 3). Subsequently, remote points (e.g., A and B) are reconnected following the shortest path in the net of nearest-neighbor distances (c). The resulting set of geodesic distances is projected to lower dimensionality (d).
2.4. Sensor simulations The ASD-spectra (i.e., 25 spectra per plot and date) were averaged to one spectrum per date and smoothed using a quadratic Savitzky–Golay filter (Savitzky and Golay, 1964) with a kernel size of 51 nm to mitigate high-frequency noise. Multispectral sensor signals were simulated based on these pre-processed spectra. These simulated sensor signals may differ from the actual sensor signals, since the simulations merely involved the spectral and radiometric resolution of the sensors; other sensor characteristics, such as spatial resolution, sensor geometry, and signal-to-noise ratio, as well as atmospheric influences were not considered. Consequently, the limitations of these simplified assumptions will be addressed in Section 4. The simulations were implemented per sensor, band, and plot following Eq. (1):
signali,x = 2rad ×
(ref × srci,x )
(1)
with signali,x the plot-based simulated signal of sensor x in band i, rad the radiometric resolution in bit, ref the plot-based canopy reflectance measured by the field spectrometer, and srci,x the spectral response curve for band i of sensor x, srci,x = 1. The band-specific spectral response curves indicate the instrument’s sensitivity towards incoming radiation. These curves were used in the simulation to calculate a weighted average of the original reflectance values. This weighted average was subsequently scaled to the respective sensor’s radiometric resolution and converted to integers. Eleven multispectral sensors were selected for comparison. Table 1 gives a brief overview of these sensors and their
characteristics. The spaceborne sensors ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer), IKONOS, Landsat 5 TM (Thematic Mapper), Landsat 7 ETM+ (Enhanced Thematic Mapper), Quickbird 2, RapidEye, SPOT 5 (Satellite Pour l’Observation de la Terre), and Worldview-2 are frequently used for remote sensing of vegetation properties. The launches of LDCM OLI and Sentinel-2, two future spaceborne sensors with promising characteristics for remote sensing of vegetation, are planned for 2013 and 2014, respectively. The HRSC-AX camera is the only airborne sensor in this comparison. Data of this or similar sensors are widely used in nature conservation since they are frequently available as aerial photographs from land survey agencies. The spectral response curves of these sensors were requested from the operating companies and institutions. The thermal bands of ASTER, Landsat 5 TM, Landsat 7 ETM+, and LDCM TIRS were not considered in this study since they are not covered by the ASD spectrometer. For better understanding why sensors differed in their performance, we considered their spectral coverage and the number of bands in the interpretation of the results. To quantify spectral coverage, we identified statistically independent spectral regions in the vegetation spectrum and analyzed whether these regions were represented by the respective sensor. To determine these regions in a data-driven way, we considered inter-correlations between neighboring wavelengths in a correlogram of the field spectrometer data. In this correlogram, local minima of correlations between neighboring wavelengths were identified and taken as boundaries to delineate the independent spectral regions. Assuming that the spectral information within these regions was highly redundant, we analyzed whether the sensors covered each region with one or more bands.
222
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
Fig. 4. Distribution of sampling dates across the vegetation period (a) and interpolated cover fractions of litter (b) and canopy height (c). NPG, nutrient-poor grassland; WH, wet heath; FPM, floodplain meadow.
2.5. Spectral separability of vegetation types The aim of this study was to track small floristic differences within each of the vegetation types based on the related spectral variability. In consequence, the floristic dissimilarity between the three vegetation types is expected to cause even larger spectral differences. A test towards spectral differences between the three vegetation types may hence help to answer the main question. Spectral data of the three vegetation types taken on a single day would have been desirable for this task. Such data was, however, not available for logistical reasons. We hence compiled three sets of data taken within a time frame of 2 weeks (Fig. 4a) and used these sets for the analysis. To test for the spectral separability of the three types, we calculated for each sensor the Jeffries–Matusita distance (JMD) between the three vegetation types. The JMD quantifies the average distance between the density functions of two compared vegetation types i and j in the multidimensional spectral feature space. It was
calculated for two types i and j following Eq. (2) (Richards and Jia, 2006): JMD = 2(1 − e−B )
(2)
where B is the Bhattacharyya distance (Eq. (3)) between two normal distributions i and j with average reflectance values mi and mj and covariance matrices ˙ i and ˙ j , respectively: B=
1 × (mi − mj )t 8
× ln
˙i + ˙j 2
|(˙i + ˙j )/2| |˙i |1/2 |˙j |1/2
−1 (mi − mj ) +
1 2
(3)
The JMD ranges from 0 for spectrally identical classes to 2 for perfect spectrally separable classes. It was calculated considering all available spectral bands of the respective sensor.
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
223
Table 1 Summary of the 11 multispectral sensors used for comparison. Bands listed in brackets were not considered in this study. Sensor
# bands
GSD at nadir (m)
Radiometric resolution (bit)
Revisiting time (days)
ASTER
3 + 1 VIS & NIR 6 SWIR (5 TIR)
15 30 (90)
8 8 (12)
16
HRSC-AX
4 VIS & NIR 1 Pan
>0.15 >0.15
12
On demand
IKONOS
4 VIS & NIR 1 Pan
3.28 0.82
11
11
Landsat 5 TM
6 VIS, NIR, & SWIR (1 TIR)
30 (120)
8
16
Landsat 7 ETM+
6 VIS, NIR, & SWIR 1 Pan (1 TIR)
30 15 (60)
8
16
LDCM OLI
8 VIS, NIR, & SWIR 1 Pan (2 TIR)
30 15 (100)
12
16
Irons et al. (2012)
Quickbird 2
4 VIS & NIR 1 Pan
2.44 0.61
11
Contents lists available at SciVerse ScienceDirect
International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag
Assessing floristic composition with multispectral sensors—A comparison based on monotemporal and multiseasonal field spectra Hannes Feilhauer a,b,c,∗ , Frank Thonfeld b , Ulrike Faude c,f , Kate S. He d , Duccio Rocchini e , Sebastian Schmidtlein b,c a
Department of Geography, University of Erlangen-Nuremberg, Kochstr. 4/4, D-91054 Erlangen, Germany Center for Remote Sensing of Land Surfaces, University of Bonn, Walter-Flex-Str. 3, D-53113 Bonn, Germany c Vegetation Geography, University of Bonn, Meckenheimer Allee 166, D-53115 Bonn, Germany d Department of Biological Sciences, Murray State University, 2112 16th Street, Murray, KY 42071, USA e Fondazione Edmund Mach, Research and Innovation Centre, Department of Biodiversity and Molecular Ecology, GIS and Remote Sensing Unit, Via E. Mach 1, 38010 S. Michele all’Adige (TN), Italy f EFTAS Fernerkundung, Oststraße 2-18, 48145 Münster, Germany b
a r t i c l e
i n f o
Article history: Received 30 July 2012 Accepted 10 September 2012 Keywords: Conservation Isometric Feature Mapping Optical remote sensing Partial Least Squares regression Subclass information Vegetation
a b s t r a c t Assessing and mapping patterns of (semi-)natural vegetation types at a large spatial scale is a difficult task. The challenge increases if the floristic variation within vegetation types (i.e., subtype variation of species composition) is the target. A desirable way to deal with this task may be to address such vegetation patterns with remote-sensing approaches. In particular data from multispectral sensors are easy to obtain, globally accessible, and often provide a high temporal resolution. They hence offer a comprehensive basis for vegetation mapping. The potential of such sensors for vegetation mapping has, however, never been thoroughly investigated. In particular, a systematic test regarding the spectral capabilities of these data for an assessment of detailed floristic variation has not been implemented to date. We thus addressed in this study the question how the ability of optical sensors to map floristic variation is affected by their respective spectral coverage and number of bands. To answer this question, we simulated monotemporal and multiseasonal data of eleven multispectral sensors. These data were used to model gradual transitions in species composition (i.e., floristic gradients) within three types of spontaneous vegetation typical for Central Europe using Partial Least Squares regression. Comparison of the model fits (ranging up to R2 = 0.76 in cross-validation) illustrated the potential of multispectral data for detailed vegetation mapping. The results show that spectral coverage of the entire solar-reflective domain is the most important sensor characteristic for a successful assessment of floristic variation. Model and sensor performances as well as limitations are thoroughly discussed, and recommendations for sensor development are made based on the final conclusions of this study. © 2012 Elsevier B.V. All rights reserved.
1. Introduction Maps of land-cover and vegetation types are crucial for decisionmaking in conservation management and planning (Leemans and Zuidema, 1995). These maps are frequently based on optical remote-sensing data acquired by air- or spaceborne sensors (Xie et al., 2008). Since satellite imagery offers the opportunity to depict the state of large areas rapidly, time series of derived mapping products may further represent a viable basis to monitor a dynamically changing landscape on a regular basis (e.g., Coops et al., 2009). Users interested in such monitoring opportunities frequently require detailed subtype information on habitat conditions (i.e., the
∗ Corresponding author. Tel.: +49 9131 85 26680. E-mail address: [email protected] (H. Feilhauer). 0303-2434/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jag.2012.09.002
fine-scale variability of a type). The requested information may include the presence of invasive species or degradation indicators, structural attributes such as shrub and grass encroachment, or the species composition of a vegetation type. In terms of rather coarse land-cover classes, these variations will be undetected due to the generalization. For conservation planning, this information may, however, be decisive for the assessment of a site in order to meet the standards of, e.g., the European Union directives on the conservation of natural habitats and of wild fauna and flora, a program aiming to protect threatened habitats and species across Europe (Council of the European Communities, 1992). In consequence, various practical considerations have been made on how remote sensing can meet this demand (see, e.g., Vanden Borre et al., 2011). Multispectral data taken by spaceborne sensors are easy to obtain for land-cover assessments of large areas. Such data have also been used to assess habitat conditions for monitoring
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
purposes (Förster et al., 2008; Franke et al., 2012; Spanhove et al., 2012; Vogelmann et al., 2012) and to detect invasive species (e.g., Dorigo et al., 2012; Ghioca-Robrecht et al., 2008); however, these assessments seldom include floristic variation within discrete habitats. Although some studies investigate spectral characteristics of specific sensors for such tasks (e.g., Schuster et al., 2012), a systematic test of the suitability of multispectral data for an assessment of detailed floristic variation has not been implemented up to now. To address this gap, we tested in this study to what extent the spectral information of multispectral data allows for a detailed assessment of the variable floristic composition within a certain natural habitat. We further tested which sensor characteristics are responsible for success or failure in this task. 2. Materials and methods 2.1. Study sites and sampling design To answer these questions, an empirical study was implemented in three different types of spontaneous vegetation very common in Central Europe: a nutrient-poor and dry grassland, a wet heath in transition to mire, and a nutrient rich and wet floodplain meadow. For the aim of this study, we tested the potential of eleven multispectral sensors to spectrally discriminate these vegetation types and to assess the within-type floristic variation with simulated data in a multivariate approach (see Fig. 1 for a flowchart showing a scheme of the study approach). We described the variability within each vegetation type as floristic gradients (i.e., gradual transitions in species composition, Gleason, 1926) since this concept is suitable to address the fuzzy characteristics of natural vegetation.
Fig. 1. Flowchart of the analysis showing logical steps of the study approach. The vegetation records sampled in each type were subjected to an ordination to quantify and describe the variability within the type as gradual transition (i.e., floristic gradient). The signals of eleven multispectral sensors were simulated based on hyperspectral data taken in the field. Monotemporal and multiseasonal simulated data were subsequently regressed against the first ordination axes. The model fits were used to evaluate the performance of the respective sensor regarding an estimation of within-type floristic variation.
219
The three sites are located near Cologne, Germany (Fig. 2a). The nutrient-poor grassland (7◦ 07 00 E, 50◦ 54 17 N, Fig. 2b) covers an area of 26 ha. It features a floristic gradient from Agrostis capillaris–Holcus mollis grassland to stands dominated by Calamagrostis epigejos. The latter result from human disturbance. The wet heath (12 ha, 7◦ 09 57 E, 50◦ 52 27 N, Fig. 2c) includes patches of heather (dominated by the dwarf shrubs Calluna vulgaris and Erica tetralix) that are interspersed along a wetness gradient with tussocks of Molinia caerulea, stands of Narthecium ossifragum, and wet hollows with Sphagnum mosses. The floodplain meadow (12 ha, 7◦ 07 27 E, 50◦ 46 34 N, Fig. 2d) is a nutrient-rich lowland meadow with gradual transitions to tall forb communities. No land use or management activities took place during the time of data acquisition. We established permanent plots for the time of data acquisition based on a random sampling design in these three sites. The plot arrangement was constrained by minimum inter-plot distances of 5 m. Further, locations in the shade of trees or shrubs were avoided. The plots were circular in shape and covered an area of 1 m2 each. Although this plot size did not meet the requirements for an image-based analysis, it was optimized for the demands of detailed vegetation sampling and the application of field spectroscopy. Each plot center was marked with a stake during the sampling period. The nutrient-poor grassland was sampled with n = 57 plots, the wet heath with n = 35 plots, and the floodplain meadow with n = 37 plots. These three sites were subsequently treated as three vegetation types (i.e., individual classes; type and class are interchangeable in this paper). 2.2. Vegetation data per type Cover fractions of all occurring vascular plants and mosses within each plot were recorded to quantify the within type floristic variation. Cover fractions were estimated close to the vegetation optimum at the respective site. The nutrient-poor grassland was sampled on July 15th, the wet heath on July 31st, and the floodplain meadow on June 23rd, 2008. The visual estimations of cover fractions were aided by a circular 1 m2 sampling frame. Variation of species composition over the vegetation period was negligible for the study sites (Feilhauer and Schmidtlein, 2011). We hence implemented all analyses using the monotemporal vegetation records. To describe the floristic variation within each vegetation type, the vegetation data of each site were separately subjected to an ordination analysis. Ordination techniques are frequently used in ecology to describe and quantify structures in vegetation through dimensionality reduction (Legendre and Legendre, 1998). Therefore, the vegetation records are arranged numerically regarding their floristic similarity in a synthetic Euclidean space. Several algorithms are available for this task. In the present study we used Isometric Feature Mapping (Isomap, Tenenbaum et al., 2000), an ordination approach with high flexibility towards data structures. Given this flexibility, Isomap has shown promising results in previous studies (e.g., Feilhauer et al., 2011; Mahecha et al., 2007). Isomap is an ordination approach that is based on a resemblance matrix of inter-plot dissimilarities in a multidimensional floristic similarity space (Fig. 3a and b). The distance metric to quantify these dissimilarities can be chosen freely. Here we used Bray–Curtis distances, a measure well established in ecological studies (Legendre and Anderson, 1999). The Isomap algorithm is based on these distances to identify k nearest neighbors for each plot. These neighborhood relations are subsequently used to generate a new resemblance matrix with geodesic distances that connect remote points via a net of mutual neighbors (Fig. 3c).
220
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
Fig. 2. Location of the three sites in Germany (a) and plot distribution across the sites (b–d). Coordinates are given in UTM 32U WGS 84. Aerial photographs in the background depict the surface in the red spectral region with a spatial resolution of 20 cm (© underlying aerial photography: state of North Rhine-Westphalia).
The geodesic distances resulting from a small set of neighbors are highly non-linear, with a growing number of considered neighbors the linearity increases. Further, the geodesic dissimilarities are projected to lower dimensionality with a multidimensional scaling approach (Fig. 3d). Prominent non-linear structures in the data can be preserved throughout the analysis by choosing an appropriate k. We used this approach to identify the main floristic gradient (i.e., the first dimension of the Isomap space) for each site. The scores of the plots on these gradients were subsequently taken as continuous description of within-type variation of the vegetation types. Indicator species for each vegetation type were identified in a numerical analysis using the isopam algorithm (Schmidtlein et al., 2010) to illustrate this floristic variation. Isomap and isopam analyses were carried out in the R statistical environment (R Development Core Team, 2011) using the packages vegan (Oksanen et al., 2011) and isopam (Schmidtlein, 2010). 2.3. Field spectroscopy The vegetation canopy of the three sites changed during the measurement period with respect to litter fractions (i.e., dead plant material, Fig. 4b), canopy height (Fig. 4c), as well as phenological stage and drought stress. Canopy reflectance of the plots was thus measured on multiple dates (n = 7 at the nutrientpoor grassland, n = 5 at the wet heath, and n = 6 at the floodplain meadow, Fig. 4a) in the 2008 vegetation period with a fullrange field spectrometer (ASD Fieldspec 3 JRTM , ASD Inc., Boulder, CO). All measurements were taken within 2 h around solar noon relative to a SpectralonTM panel (Labsphere Inc., North Sutton, NH). These reflectance spectra covered the spectral range of 350–2500 nm in 2151 bands. The spectrometer probe was
leveled in nadir approximately 110 cm above the canopy. With the given field-of-view of 25◦ , five measurements were required to cover the whole plot area. Each measurement was taken in five repetitions. Thus, for every plot a total of 5 × 5 spectra was taken per date. We compared eleven sensors (described in Section 2.4 and Table 1) with monotemporal (i.e., one single date) and multiseasonal (i.e., multiple dates of different seasons but within a single year) data to test their performance. In practice, monotemporal analyses offer the advantage of a reduced effort and lower costs for data acquisition and preprocessing. Optimal timing of the data acquisition is crucial because the strength of statistical relations between vegetation patterns and reflectance is temporally variable. Many studies thus choose a distinct phenological stage for data acquisition (e.g., Andrew and Ustin, 2008; Ge et al., 2006; Laba et al., 2005). In an earlier study (Feilhauer and Schmidtlein, 2011), it was shown for hyperspectral data that the vegetation patterns within the sites under investigation could be most accurately differentiated during the vegetation optimum, which was defined as the date of maximum cover of photosynthetically active vegetation (i.e., the growth peak). We hence used this date for the monotemporal comparison (Fig. 4a). Yet, the configuration of various multispectral sensors (e.g., RapidEye, Sentinel-2, Worldview-2) is optimized towards a short revisit time. These sensors are predisposed to provide multiseasonal data covering phenological events continuously rather than providing a single snapshot in time. We thus quantified the additional explanatory power gained by the use of multiseasonal data considering all available time steps (i.e., the ASD sampling dates in 2008). Monotemporal and multiseasonal results gained by different sensors were compared to assess the explanatory power regarding within-type variation of vegetation.
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
221
Fig. 3. Schematic illustration of the Isomap approach. From the distribution of sample points in an n-dimensional similarity space (a), mutual inter-point distances are calculated (b). This set is used to identify the k nearest neighbors of each point (c, shown for k = 3). Subsequently, remote points (e.g., A and B) are reconnected following the shortest path in the net of nearest-neighbor distances (c). The resulting set of geodesic distances is projected to lower dimensionality (d).
2.4. Sensor simulations The ASD-spectra (i.e., 25 spectra per plot and date) were averaged to one spectrum per date and smoothed using a quadratic Savitzky–Golay filter (Savitzky and Golay, 1964) with a kernel size of 51 nm to mitigate high-frequency noise. Multispectral sensor signals were simulated based on these pre-processed spectra. These simulated sensor signals may differ from the actual sensor signals, since the simulations merely involved the spectral and radiometric resolution of the sensors; other sensor characteristics, such as spatial resolution, sensor geometry, and signal-to-noise ratio, as well as atmospheric influences were not considered. Consequently, the limitations of these simplified assumptions will be addressed in Section 4. The simulations were implemented per sensor, band, and plot following Eq. (1):
signali,x = 2rad ×
(ref × srci,x )
(1)
with signali,x the plot-based simulated signal of sensor x in band i, rad the radiometric resolution in bit, ref the plot-based canopy reflectance measured by the field spectrometer, and srci,x the spectral response curve for band i of sensor x, srci,x = 1. The band-specific spectral response curves indicate the instrument’s sensitivity towards incoming radiation. These curves were used in the simulation to calculate a weighted average of the original reflectance values. This weighted average was subsequently scaled to the respective sensor’s radiometric resolution and converted to integers. Eleven multispectral sensors were selected for comparison. Table 1 gives a brief overview of these sensors and their
characteristics. The spaceborne sensors ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer), IKONOS, Landsat 5 TM (Thematic Mapper), Landsat 7 ETM+ (Enhanced Thematic Mapper), Quickbird 2, RapidEye, SPOT 5 (Satellite Pour l’Observation de la Terre), and Worldview-2 are frequently used for remote sensing of vegetation properties. The launches of LDCM OLI and Sentinel-2, two future spaceborne sensors with promising characteristics for remote sensing of vegetation, are planned for 2013 and 2014, respectively. The HRSC-AX camera is the only airborne sensor in this comparison. Data of this or similar sensors are widely used in nature conservation since they are frequently available as aerial photographs from land survey agencies. The spectral response curves of these sensors were requested from the operating companies and institutions. The thermal bands of ASTER, Landsat 5 TM, Landsat 7 ETM+, and LDCM TIRS were not considered in this study since they are not covered by the ASD spectrometer. For better understanding why sensors differed in their performance, we considered their spectral coverage and the number of bands in the interpretation of the results. To quantify spectral coverage, we identified statistically independent spectral regions in the vegetation spectrum and analyzed whether these regions were represented by the respective sensor. To determine these regions in a data-driven way, we considered inter-correlations between neighboring wavelengths in a correlogram of the field spectrometer data. In this correlogram, local minima of correlations between neighboring wavelengths were identified and taken as boundaries to delineate the independent spectral regions. Assuming that the spectral information within these regions was highly redundant, we analyzed whether the sensors covered each region with one or more bands.
222
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
Fig. 4. Distribution of sampling dates across the vegetation period (a) and interpolated cover fractions of litter (b) and canopy height (c). NPG, nutrient-poor grassland; WH, wet heath; FPM, floodplain meadow.
2.5. Spectral separability of vegetation types The aim of this study was to track small floristic differences within each of the vegetation types based on the related spectral variability. In consequence, the floristic dissimilarity between the three vegetation types is expected to cause even larger spectral differences. A test towards spectral differences between the three vegetation types may hence help to answer the main question. Spectral data of the three vegetation types taken on a single day would have been desirable for this task. Such data was, however, not available for logistical reasons. We hence compiled three sets of data taken within a time frame of 2 weeks (Fig. 4a) and used these sets for the analysis. To test for the spectral separability of the three types, we calculated for each sensor the Jeffries–Matusita distance (JMD) between the three vegetation types. The JMD quantifies the average distance between the density functions of two compared vegetation types i and j in the multidimensional spectral feature space. It was
calculated for two types i and j following Eq. (2) (Richards and Jia, 2006): JMD = 2(1 − e−B )
(2)
where B is the Bhattacharyya distance (Eq. (3)) between two normal distributions i and j with average reflectance values mi and mj and covariance matrices ˙ i and ˙ j , respectively: B=
1 × (mi − mj )t 8
× ln
˙i + ˙j 2
|(˙i + ˙j )/2| |˙i |1/2 |˙j |1/2
−1 (mi − mj ) +
1 2
(3)
The JMD ranges from 0 for spectrally identical classes to 2 for perfect spectrally separable classes. It was calculated considering all available spectral bands of the respective sensor.
H. Feilhauer et al. / International Journal of Applied Earth Observation and Geoinformation 21 (2013) 218–229
223
Table 1 Summary of the 11 multispectral sensors used for comparison. Bands listed in brackets were not considered in this study. Sensor
# bands
GSD at nadir (m)
Radiometric resolution (bit)
Revisiting time (days)
ASTER
3 + 1 VIS & NIR 6 SWIR (5 TIR)
15 30 (90)
8 8 (12)
16
HRSC-AX
4 VIS & NIR 1 Pan
>0.15 >0.15
12
On demand
IKONOS
4 VIS & NIR 1 Pan
3.28 0.82
11
11
Landsat 5 TM
6 VIS, NIR, & SWIR (1 TIR)
30 (120)
8
16
Landsat 7 ETM+
6 VIS, NIR, & SWIR 1 Pan (1 TIR)
30 15 (60)
8
16
LDCM OLI
8 VIS, NIR, & SWIR 1 Pan (2 TIR)
30 15 (100)
12
16
Irons et al. (2012)
Quickbird 2
4 VIS & NIR 1 Pan
2.44 0.61
11
