Ruiliang Pu and Peng Gong
I. INTRODUCTION
- Both biophysical and biochemical parameters, such as LAI, species, nitrogen content, are important variables for quantifying the energy and mass exchange characteristics of a terrestrial ecosystem.
- Use of traditional multispectral data (such as MSS, TM) is limited by the amount of spectral details.
- Hyperspectral data are very useful for estimating these parameters. Such data have dozens to hundreds of consecutive narrow spectral bands.
- Hyperspectral data characterize spectral signatures with sufficient spectral sampling to allow the unambiguous identification of certain ground targets and the quantitative analysis of subtle spectral change.
II. SPECTRAL REFLECTANCE COLLECTION
- Figure 1 shows that a spectral measurement is being taken.
- Spectrometer: PSD1000 used for taking measurements in the field. The spectral resolution is approximately 2.6 nm. It is controlled by a portable computer.
- Study site: Blodgett Forest Research Station, the University of California, Berkeley.
- Conifer species: Douglas fir (DF), Giant sequoia (GS), Incense cedar (IC), Ponderosa pine (PP), Sugar pine (SP), and White fir (WF).
Collect spectral measurements (Figure 1)
III. SPECTRA OF SOME CONIFER SPECIES
- Figure 2 presents average spectral reflectances of six conifer species: DF, GS, IC, PP, SP, and WF.
Transformation (Figure 2)
IV. ANALYSIS METHODS
- Derivative of original spectra (R), D(R),
- Logarithm of R, LOG(R),
- Normalized R, N(R),
- Derivative of LOG(R), D(LOG(R)),
- Logarithm of N(R), LOG(N(R)), and
- Derivative of N(R), D(N(R)).
Band aggregation
Recognition method: Neural network algorithm
V. RESULTS
Figure 4a. reflectance difference caused by illumination change from same tree species (DF, GS, and IC). Sunlit = spectral reflectance measured at sunlit position of canopies; shaded = measured at shaded position of canopies. Figure 4b. The 1st order derivative derived from Figure 4a. Similar derivative spectra show partial removal of effects due to light changes on target spectra for each species.
Figure 3. Summary of species recognition accuracies of six transformation types with an artificial neural network algorithm.
Figure 5a. Species recognition accuracies obtained from the Hyperspectral data as a function of band width (47-54 nm). Figure 5b obtained from derivative Hyperspectral data as a function of band width (about 20 nm).
Recognition accuracies (Figure 3)
Reflectance difference (Figure 4a)
Derivative of Figure 4a (Figure 4b)
Effect of band width, original spectra (Figure 5a)
Effect of band width, derivative spectra (Figure 5b)
VI. CONCLUSIONS
Hyperspectral data can be used to recognize conifer species that have similar spectral signatures with high accuracy.
Derivative is an effective technique of transforming Hyperspectral data.
Neural network algorithm is a robust tool to identify conifer species with Hyperspectral data.
A band width of 20 nm or narrower is recommended for the recognition of the six conifer species.