Parallel Coordinate Visualization

A parallel coordinate visualization of the mean and standard deviations of each class can be used to identify differences and overlaps in the spectra of the tree species. Fig. 4.9 shows the spectra of the six species in the test area Arnsberg that are represented by a sufficient amount of sample data points. Those species are the four most important tree species groups in NRW, namely oak, beech, other broadleaved species with short rotation time (OBS.) and spruce, and in addition larch and Douglas fir. The according graphs for all nine species groups including the underrepresented species can be found in section B.1.

The first section of the parallel coordinate visualization shows the mean values and standard deviations in the bands that were extracted from the airborne image data (blue (B), green (G), red (R), near infrared (NIR)). The subsequent four char- acteristics (NIR, R, G, short wavelength infrared (SWIR)) were extracted from the SPOT satellite data. The next five bands (B, G, R, red edge (RE), NIR) contain the features of the RapidEye satellite data and the last value (intensity (I)) is extracted from the LIDAR intensity data.

In Fig. 4.9 the RapidEye bands were scaled to values in the range of the other bands, which is [0 . . . 255]. It is interesting to note, that the interspecies variability

B G R NIR NIR R G SWIR B G R RE NIR I 20 40 60 80 100 120 140 160

Airborne: B, G, R, NIR bands; SPOT: NIR, R, G, SWIR bands; RapidEye: B, G, R, RE, NIR bands; LiDAR Intensity band

Reflectances oak beech obs. larch spruce Douglas fir

Figure 4.9: Parallel coordinate visualization - Mean values and standard deviation: The first 4 bands (B, G, R, NIR) are from the airborne image, the next 4 bands (NIR, R, G, SWIR) are SPOT satellite bands, the following 5 bands (B, G, R, RE,

NIR) are RapidEye satellite bands and the last band (I) is the intensity from the airborne LiDAR data.

in all infrared bands, including the SWIR and I bands, are higher than in the visible bands. Unfortunately, the intraspecies variability is also higher for most species. Although three NIR bands are available, one from the airborne images, one from the SPOT satellite and one from the RapidEye satellite, the characteristics differ to some extent. OBS. has a higher mean value than oak in the airborne and RapidEye NIR bands, but approximately the same mean value in the SPOT NIR band. Larch has a lower mean value in the RapidEye NIR band than any other species whereas in the airborne and SPOT NIR bands, the value is higher than the mean values of spruce and Douglas fir. Larch has also different characteristics in the two SWIR bands, which are the SPOT SWIR band and the LIDAR I band. While in the SPOT SWIR band, the mean value of larch is between the two coniferous species (spruce and Douglas fir) and the three deciduous species (oak, beech and OBS.), in the I band it almost reaches the mean value of beech, which has the highest mean value in this band. Factors that contribute to the different characteristics in similar spectral bands of different sensors are the acquisition time and date, which includes weather conditions during and before the acquisition date and foliation, and sensor specifications.

Based on Fig. 4.9, the findings given in [49] that the blue band is the best indi- vidual band for tree species discrimination, cannot be confirmed on this data set. The interspecies variability in the B band is lower than in any other band of the same data source, while the intraspecies variability reaches almost the same values as in the other bands. Fig. 4.9 also illustrates that the spectral characteristics of different tree species have low interspecies variability and high intraspecies variance. The variability of the spectral reflectance is influenced by soil, site and lighting con- ditions. Fig. 4.9 also shows that e.g. beech and other broadleaved species with long rotation time (OBL.) have significantly higher mean values in the airborne infrared band, especially relative to the mean values of the blue, green and red bands. To take a closer look at the relative changes in the spectra, difference and ratio bands were calculated. For the airborne images, all possible band combinations were calculated. From the additional satellite data sources, only the band combinations involving a band that is not part of the airborne data set (SPOT SWIR and RapidEye RE) are calculated. The band combinations are shown in Fig. 4.10 and Fig. 4.11 respectively. Difference bands have also been stated in [52] to have the ability to decrease the variability in the spectral reflectance, which can occur due to environmental factors as lighting changes or changes in the water content of the leaves.

NIR−R NIR−G NIR−B B−G B−R G−R SWIR−NIR SWIR−R SWIR−G RE−B RE−G RE−R NIR−RE −60 −40 −20 0 20 40 60 80 Difference Bands Reflectances oak beech obs. larch spruce Douglas fir

Figure 4.10: Difference Bands

For the difference and ratio bands, the median will be used instead of the mean value, and the standard deviation is replaced by the 0.1587 and 0.8413 quantiles. The range that is depicted by these quantiles contains 68.3% of the data points,

as does the standard deviation for a Gaussian distribution. The difference bands are calculated by subtracting the values of two bands for each sample. Thereby, N IR − R describes a feature that is calculated by subtracting the value of the R band from the value of the NIR band.

The difference bands in Fig. 4.10 show some interesting and distinctive properties. The interspecies variability of spruce and Douglas fir on one side and oak, beech, OBS. and larch on the other side is high in the SW IR−R band while the intraspecies variabilities are rather low. A similar effect can be observed in the SW IR − G band and in the difference bands between the NIR and the visible bands for the coniferous species on one hand and the broadleaved species on the other hand. The mean value of the OBS. species differs a bit from the remaining broadleaved species in the NIR − RE band although intraspecies variabilities are quite large.

The ratio bands in Fig. 4.11 are calculated by dividing the mean value of one band by the mean value of another band for each object. Therefore, the feature N IR/Ris calculated by dividing the mean value of the infrared band by the mean value of the red band.

NIR/R NIR/G NIR/B B/G B/R G/R SWIR/NIR SWIR/R SWIR/G RE/B RE/G RE/R NIR/RE 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 Ratio Bands Reflectances oak beech obs. larch Spruce Douglas fir

Figure 4.11: Ratio Bands

The characteristics of the first six ratio bands are rather similar to the charac- teristics of the first six difference bands. In the band combinations of the NIR and the visible airborne bands, the interspecies variance between oak, beech and OBS. in one group and spruce, larch and Douglas fir in the other group is high. There is also a difference in the band calculated from the SWIR and the R band. The

relation between interspecies and intraspecies variability is more advantageous in the according difference band than in the ratio band. However, the same can be observed for the band calculated form SWIR and G.

4.2.1.1 Hyperspectral Data

For a small part of the test area Arnsberg, a hyperspectral data set was available. This data set was explored to find wavelength ranges that are interesting for tree species classification. Hyperspectral data still is either too expensive or has a reso- lution that is too low for applications on the single tree level. The analysis aimed at finding suitable bands to choose appropriate satellites and sensors for practical very high resolution applications.

Fig. 4.12 shows the entire available hyperspectral spectrum of the Airborne Imag- ing Spectroradiometer for Applications (AISA) sensor ranging from 0.975 µm to 2.449 µm with a band width of 6.3 nm for each band. It also shows that there are

1 1.2 1.4 1.6 1.8 2 2.2 2.4 0 20 40 60 80 100 120 140 160

Mean Values of AISA

AISA Band frequencies

Reflectances oak beech obs. larch spruce Douglas fir

Figure 4.12: Hyperspectral spectrum. The wavelengths are given in [nm]. three interesting regions with different reflectance values. The first region ranges from 0.975 µm to about 1.4 µm, which is the area that is often referred to as NIR although NIR is defined as the spectral range from 0.78 µm to 3.0 µm by [29] where the range from 0.78 µm to 1.4 µm is defined as IR-A. The range, that can be cap- tured by photographic film is part of the IR-A and ranges from 0.7 µm to 1.0 µm. Glscir usually refers to photographic infrared within this range. The margin in the definition of the NIR between short wavelength range of the near infrared (IR-A)

and long wavelength range of the near infrared (IR-B) at 1.4 µm is defined due to the increased water absorption at this particular wavelength. The second region of interest of the AISA spectrum ranges from about 1.4 µm to 1.9 µm. The margin at 1.9 µm is a second area with increased water absorption, which again leads to very little reflectance in the spectral bands around this wavelength. The third region of interest in this spectrum ranges from 1.9 µm to 2.4 µm. The spectra of the three parts are shown in Fig. 4.13.

In the first part of the hyperspectral spectrum in Fig. 4.13a, the sequence of the species stays the same. Although the interspecies variance seems distinct, the high intraspecies variances make it difficult to find features that allow a clear discrim- ination between the species. In the second part of the spectrum, between 1.4 µm and 1.9 µm, the interspecies variance between spruce and Douglas fir declines. The variance between the characteristics of Douglas fir and larch increases in the second part of the hyperspectral spectrum compared to the first part. The third part of the hyperspectral range is between 1.9 µm and 2.4 µm and shown in Fig. 4.13c. In this part spruce and Douglas fir are even more similar and the interspecies variance between larch and oak also declines further. The figures without standard deviations for all nine species are given in section B.2. For the calculation of the difference and ratio bands, 22 bands located at local maxima and minima were selected. These bands are given in table 4.2.

Table 4.2: Selected hyperspectral bands

Band Wavelength Band Wavelength Band Wavelength

10 1.032 µm 99 1.592 µm 185 2.134 µm 25 1.126 µm 100 1.599 µm 191 2.172 µm 42 1.233 µm 104 1.624 µm 194 2.190 µm 47 1.265 µm 106 1.636 µm 195 2.197 µm 50 1.284 µm 108 1.649 µm 199 2.222 µm 69 1.403 µm 146 1.888 µm 234 2.442 µm 91 1.542 µm 151 1.920 µm 95 1.567 µm 154 1.939 µm

The difference bands are calculated from the bands in table 4.2 by taking the first band (band 10 of the AISA spectrum) and subtracting each of the remaining bands

1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 0 20 40 60 80 100 120 140 160 180 200

Mean Values of AISA − first part of the spectrum − with std−bars

AISA Band frequencies

Reflectances oak beech obs. larch spruce Douglas fir

(a) First part

1.4 1.45 1.5 1.55 1.6 1.65 1.7 1.75 1.8 0 5 10 15 20 25 30

Mean Values of AISA − second part of the spectrum − with std−bars

AISA Band frequencies

Reflectances oak beech obs. larch spruce Douglas fir (b) Second part 1.95 2 2.05 2.1 2.15 2.2 2.25 2.3 2.35 2.4 2.45 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Mean Values of AISA − third part of the spectrum − with std−bars

AISA Band frequencies

Reflectances oak beech obs. larch spruce Douglas fir (c) Third part

one after another. Then the second band (band 25) is taken and each subsequent band (bands 42 to 234) is subtracted and so on. The calculated bands are numbered in ascending order. As shown in the Fig. 4.14a, the characteristics of the difference bands are, apart from a scaling factor, similar in most regions.

20 40 60 80 100 120 140 160 180 200 220 −50 0 50 100 150

Difference bands AISA

Difference Bands Values oak beech obs. larch spruce Douglas fir

(a) Difference bands calculated from the selected hyperspectral bands

0 5 10 15 20 20 40 60 80 100 120 140 160 180 200

Difference bands AISA with errorbars − part 1

Difference Bands Values 25 26 27 28 29 30 31 32 33 2 4 6 8 10 12 14

Difference bands AISA with errorbars − part 2

Difference Bands

Values

(b) Subsets of the difference bands. Calculated from the 10th and all selected bands and from the 25th and the bands between the 91st and the 108th respectively.

Figure 4.14: Difference bands

Two details of the difference bands are shown in Fig. 4.14b. The first image shows the difference bands calculated as the differences between the 10th band and the remaining selected bands. It is interesting to note, that the proportion of interspecies and intraspecies variability is more advantageous than that of the difference bands calculated only from visible and red edge bands. The second image in Fig. 4.14b shows the difference bands calculated from the 25th band and the

selected bands between the 91st and the 108th. In these bands, the mean value of OBS. is higher than the mean value of beech. The mean value of larch is below the mean value of spruce in the last five bands. The mean value of beech even drops below the mean value of oak in the difference bands 30 and 31, which are calculated as band 25 minus band 104 and 106 respectively. Unfortunately the relation between interspecies and intraspecies variance is less advantageous than in the first 21 difference bands.

20 40 60 80 100 120 140 160 180 200 220 0 10 20 30 40 50 60 70 80

Ratio bands AISA

Ratio Bands Values oak beech obs. larch spruce Douglas fir

(a) Ratio bands calculated from local minima and maxima in the spectrum

16 17 18 19 20 3.5 4 4.5 5 5.5 6 6.5 7 7.5

Ratio bands AISA with errorbars − part

Ratio Bands Values oak beech obs. larch spruce Douglas fir

(b) Subset of the ratio bands. Calculated from the 10th and the bands between the 185th and the 199th respectively.

Figure 4.15: Ratio bands

similar to the difference bands and are shown in Fig. 4.15a. As the ratio bands are calculated from maximum and minimum bands, they have very different scales as the ratio between two maxima will be much smaller than a maximum divided by a band in a local minimum. The latter will also always have very large variances, as only small changes in the value of the minimum band will lead to large changes in the ratio band values. Fig. 4.15b shows a small part of the ratio bands. These are calculated from the 10th and the bands between the 185th and 199th band. Therefore, they combine the NIR region and the SWIR region. In these bands, spruce has a similar mean value to oak. Larch has a good separability from Douglas fir due to a interspecies variability that has a similar magnitude as the sum of the two intraspecies variabilities.

The figures above show that some species are especially hard to separate, e.g. OBS. and beech. Furthermore, the characteristics of the bands are quite similar within each spectral regions such that additional spectral bands that cover these spectral regions can help classifying tree species, but the additional benefit of the high spectral resolution is unlikely to live up to the expectations for most of the bands. The according figures containing all available tree species, including the species that are underrepresented in the reference data set, can be found in sec- tion B.2.