Home Articles Improving Species Spectral Discrimination Using Derivatives Spectra For Mapping of Tropical Forest...

Improving Species Spectral Discrimination Using Derivatives Spectra For Mapping of Tropical Forest From Airborne Hyperspectral Imagery


Affendi Bin Suhaili
PhD Candidate [GS14198]
Faculty of Forestry
Universiti Putra Malaysia
UPM Serdang, Selangor.
[email protected]

Helmi Z.M Shafri

N.A Ainuddin

A.G. Awg Noor

I.Faridah Hanum

ABSTRACT
Forest resource maps were traditionally prepared from forest inventories involving aerial photography and fieldwork, however with the advent of technology, remote sensing from satellite platforms offers an alternative and economic tool for forest mapping. With the development of hyperspectral sensors, forest species discrimination and mapping could be improved as the fine spectral resolutions inherent in this system allows the identification of small differences in the similar spectral responses between forest species. There are however some problems associated with the use of this technology, namely associated with the high dimensionality of the data set that results in redundancy and the also the detection of small absorption features presence in the plant spectra. By using derivative spectra, subtle spectral features between different tree species could be detected from a limited number of bands and the overlapping absorption characteristics from these similar plant responses are resolved which might not be possible from analysis of the original spectrum. This study evaluates the discriminative capabilities in terms of the spectral separability among tropical tree species and on classification accuracy when using the derivative spectra for mapping an old growth forest plot in the Forest Research Institute of Malaysia (FRIM), Kepong, Selangor. Results showed that the separability (JM distance measure) of crown spectra from 16 species of trees commonly found in the Malaysian tropical forest were higher and classification accuracy (Max-Likelihood algorithm) improved to 70.15% when using the derivative data set. From this study it is concluded that hyperspectral data obtained from the airborne imaging spectrometer (AISA) could discriminate between tree species of the tropical forest and the use of derivative spectroscopy as an image enhancement technique could improve species discrimination and classification accuracy. This has provided a potential for application of the hyperspectral imaging system for mapping of the tropical forest at an operational level.

INTRODUCTION
Forests have long been regarded as a national treasure in Malaysia. With the current depletion of forested areas around the world, it is important that we manage these renewable resources in a sustainable manner and in order to formulate and excise efficient forest management policies and practices, it is important to have the maximum information about the forest cover. Forest resource maps were traditionally prepared from forest inventories involving aerial photography and fieldwork, however with the advent of technology, remote sensing from satellite platforms offers an alternative and economic tool for forest mapping. The tropical rainforest is much known for its high species composition and the current use of broad band multispectral sensors would not be effective to distinguish the small spectral differences of the forest canopies due to the similar spectral signature between the tree species.

With the development of hyperspectral sensors, forest species discrimination and mapping could be improved as the fine spectral resolutions inherent in this system allows the identification of small differences in the similar spectral responses between forest species. There are however some problems associated with the use of this technology, namely associated with the high dimensionality of the data set that results in redundancy and the also the detection of small absorption features presence in the plant spectra. Operational application of the hyperspectral sensors for mapping of forest canopies are also subjected to varying illumination conditions and background effects. One method which is commonly employed to resolve or enhance the absorption features that might be masked by interfering background absorption is by the use of derivative spectrometry (Curran et al.,1990 ; Filella and Penuelas, 1994). Spectral derivatives also aid in suppressing the continuum caused by other leaf biochemicals (such as lignin and secondary pigments) and canopy background effects (Elvidge, 1990). By using derivative spectra, subtle spectral features between different tree species could be detected from a limited number of bands and the overlapping absorption characteristics from these similar plant responses are resolved which might not be possible from analysis of the original spectrum.

   

MATERIALS AND METHODS

Study Site
This study was conducted at an old growth plot (as early as 1927) in the Forest Research Institute of Malaysia (FRIM) Kepong, Selangor. The reserve is situated 11km north-west of Kuala Lumpur with the latitude 3013′ to 3015’N and longitude of 101036′ and 101039′ E. The study was confined to a mixed species plot, which comprise 16 tree species from both Dipterocarps (Meranti Sengkawang Merah, Meranti Paang, Meranti Tembaga, Meranti Rambai Daun, Kapur, Keladan, Balau Laut, Balau Kumus and Merawan Siput Jantan) and Non-Dipterocarps (Jelutong, Kelat, Pulai, Karas, Sesendok, Melembu and Inggir Burong) taxa.

Hyperspectral Imagery
The airborne campaign over FRIM was conducted on May 2005 using a compact push-broom airborne imaging spectrometer (AISA), covering the 430nm to 900nm. The sensor is integrated with a GPS/IMU system which measures the position, velocity, timing altitude and linear accelerations of the sensor head during flight. The system also consists of an internal fiber optic downwelling irradiance sensor (FODIS), which provide information of radiation during the measurement as it comes from the source, the sun, and use that information to correct the data imaged from the ground to at sensor reflectance. As the objective of the study was to map individual tree species, the AISA sensor was operated at spatial mode with a 1m ground pixel and 20 wavelength channels spectrally configured (Affendi et al. 2005) for tropical forest mapping.

First Derivative Reflectance
Derivative spectroscopy is applied by numerically computing the derivatives of the original spectrum with respect to the wavelength and using them to detect the absorption band positions of the spectrum. In this study, the 1st derivative transformation is calculated using IDL’s DERIV function (David Gorodetzky) a plug-in to the ENVI 4.0 software. It performs a numerical differentiation using 3-point Lagrangian interpolation where the wavelength of the respective spectral bands (AISA) are used as input for defining the spacing between the bands.

Separability and Accuracy Measures
A comparison was made to the Reflectance data set and its 1st Derivative Transformation in order to evaluate the effectiveness of using this technique in spectrally discriminating between the tree species. A measure of separation of the various tree species based on their spectral response (pixel scale) was obtained with the Jefferies-Matuista distance measure and the classification accuracy of the data sets were assessed based on an error or confusion matrix and the kappa coefficient which compares the crown region of interest (ROI) from the field measured data to the tree crown polygons from the respective data sets established by the Maximum Likelihood algorithm.

RESULTS AND DISCUSSION
Visual assessment of the image shows that the 1st derivative data set could better discriminate among the different tree species classes in the image based on the distinct tonal variations present as compared to the reflectance data set. Figure 1 compares a false color composite image from the original reflectance data set (a) and the spectrally enhanced 1st derivative data set (b) using band 13(R), 11(G) and 9(B) in the mixed forest plot.


Figure 1a: False composite image [R(Band 13) G(Band 11) B(Band 9)] of the AISA reflectance data set over the old growth forest plot.


Figure 1b: False composite image [R(Band 13) G(Band 11) B(Band 9)] of the transformed 1st derivative data set over the old growth forest plot.

As shown in Figure 2, although the original reflectance spectra have almost identical values between 680nm and 730nm (red-edge) region, they differ in magnitude and to some extent shapes in the 1st derivative spectra over the same wavelength range. Several interesting spectral features are apparent in the derivative spectra that were obscure in the original spectra. For example is the double-peak feature that is observed on the canopy derivative reflectance, which according to the study by Zarco-Tejada et al.(2003) is a function of the steady-state natural fluorescence emission bands centered at 690nm and 730nm. Horler et al. (1983), on the other hand attributed the first peak at around 700nm to chlorophyll content in the plant leaves and the second at around 725nm to cellular scattering in the leaf. Based on these features that were observed in the derivative data set of our study, we were able to discriminate between the dipterocarp and non dipterocarp genera. This could be seen from the spectral derivative plots in Figure 2(bottom), where the dipterocarps has a higher 2nd peak (centred at 730nm) as compared to the non dipterocarps which generally has an equal or higher 1st peak (centred at 690nm).


Figure 2: Double peak feature as apparent in the spectrally enhanced red edge region of the 1st derivative spectra (bottom) as compared to the reflectance spectra

Spectral Separability
The dipterocarps are slightly more separable as compared to the non dipterocarps. This could be seen from the results which show distinct separation of 8 species (Meranti Sengkawang Merah, Meranti Paang, Meranti Tembaga, Meranti Rambai Daun, Kapur, Balau Laut, Balau Kumus and Merawan Siput Jantan) as compared to 6 species (Sesendok, Inggir Burong, Karas, Pulai, Kelat and 2 classes of the Jelutong’s) of non dipterocarps from the derivative image data set. This result also shows that the Shorea’s (Meranti’s and Balau) are spectrally distinct from the other species groups. 65.8% of the spectral derivatives data set shows full separation (2.0) based on the JM distance measure with an overall improvement of 95.3% as compared to the reflectance data set. Based on the reflectance spectra, the Keladan (D.oblongfolia) groups (KLD1 and KLD2) have similar response to the other 14 species, however a 50% improvement in separability could be seen for KLD1 using the spectral derivatives data. The derivatives spectra seems to improve the separability of the non dipterocarp groups from only 2 species (Sesendok and Inggir Burong) which are distinctly separable in the reflectance data set to 6 species (Sesendok, Inggir Burong, Karas, Pulai, Kelat and Jelutong) when using the derivatives data set.

Classification Accuracy
An error matrix was calculated by comparing the crown ROI’s from the field data to the tree crown polygons from the respective data sets established by the Maximum Likelihood algorithm. The overall accuracy of the classification from the reflectance data set was 67.58%, with 7247 out of 10723 pixels correctly classified, with a kappa value of 0.64. Improvement in accuracy is seen in the spectral derivative data set with an overall accuracy of 70.15%, with a kappa value of 0.67 (Figure 3). Only 2501 pixels were misclassified which shows an improvement of 13.5% over the reflectance data set. Highest misclassifications were among Inggir Burong (IBU2), Meranti Paang (MPA) and Jelutong (JEL3) on both data sets however there are improvements in accuracy in the derivative data set with the exception of the Jelutong (JEL3) class.


Figure 3: Classified image of 1st derivative data set using the maximum likelihood classifier

CONCLUSION
This study has shown that the use of derivative analysis can improve the separability of canopy spectra of the tree species, even if there are spectrally similar, which is common to tropical forest environment. It has shown that the subtle differences within the red edge spectral region could be enhanced based on both amplitude and shape differences using the 1st derivative spectra hence giving the characteristic signature of the respective tree species. From this study it is concluded that hyperspectral data obtained from the airborne imaging spectrometer (AISA) could to a certain extent discriminate between tree species of the tropical forest and the use of derivative spectroscopy as an image enhancement technique could improve species discrimination and classification accuracy. This has provided a potential for application of the hyperspectral imaging system for mapping of the tropical forest at an operational level.

ACKNOWLEDGEMENT
The authors thank Datuk Amar Haji Abdul Aziz Dato Haji Husain, the State Secretary of Sarawak for the support and encouragement, financial support from the Sarawak State Government (UPM vote 62188), Dr. Noor Azlin Yahya of FRIM for providing the research facilities and field support. Special thanks to Jukka Okkonen, Specim Finland for the technical advice and Aeroscan Precision (M) Sdn.Bhd for the airborne data collection.

REFERENCES

  • Affendi Suhaili, Ainuddin, N.A., Awg Noor, A.G., Faridah Hanum, I., Shafri, H.Z.M. and Manohar Mariapan, 2005. Optimized selection of hyperspectral bandsets for mapping of tropical timber trees using hyperspectral imaging sensors. 14th Malaysian Forestry Conference, Kota Kinabalu, Sabah, Sept. 12-16. 9p.
  • Curran P.J., Dungan, J.L. and Gholz, H.L. 1990. Exploring the relationship between reflectance red edge and chlorophyll content in slash pine. Tree Physiology, 7: 33-48.
  • Elvidge, C.D, Chen, Z. and Groeneveld, D.P. 1993. Detection of trace quantities of green vegetation in 1990 AVIRIS data. Remote Sensing of Environment, 44: 271-279.
  • Filella, I. and Penuelas, J. 1994. The red edge position and shape as indicators of plant chlorophyll content, biomass and hydric status. International Journal of Remote Sensing, 15(7) 1459-1470.
  • Horler, D.N.H., Dockray, M. and barber, J., 1983. The red edge of plant leaf reflectance. International Journal of Remote Sensing, 4(2) 273-288.
  • Zarco-Tejada, P.J., Pushnik, J.C, Dobrowski and Ustin, S.L., 2003. Steady state chlorophyll a fluorescence detection from canopy derivative reflectance and double-peak red-edge effects. Remote Sensing of Environment, 84: 283-294.