Classification of algal bloom types
From remote sensing reflectance
3. Spectral Reflectance Signatures Of Algal Bloom Classes
Spectral reflectance refers to the ratio of the detected radiance reflected from a target surface to the total incidence irradiance. In this project, a handheld spectrometer (GER 1500) was used to measure the radiance reflected from the sea surface. The detected radiance from the sea surface was normalized by the radiance reflected off the surface of a reference white plate to obtain the reflectance of the sea surface. The spectrometer has 512 wavelength channels covering the wavelength from 350 nm to 1050 nm, with a wavelength resolution of 2 nm.
From the collected spectra, SeaWiFS and MERIS data are simulated according to the band specifications shown in Table 2. The simulation is done by integrating the spectrometer radiance within each specified wavelength window to obtain the desired radiance for the corresponding SeaWiFS and MERIS channels. A flat spectral response curve is assumed for each of the satellite sensor channels. Only the channels in the visible region (400 nm to 760 nm) are considered in the simulation. Hence, the first 6 bands of the SeaWiFS sensor and the first 10 bands of the MERIS sensor are simulated. The simulated SeaWiFS and MERIS spectral reflectance data for the reference sea water and the eight algal bloom classes are shown in Fig. 1. Each spectrum shown in Fig. 1 is the mean of a set of spectra corresponding to the reference sea water and each of the algal bloom classes. The spectra have been normalized so that each of them has a mean value of zero and a variance of one. In this way, the magnitude of reflectance has no influence on the normalized spectra, and the shapes of the spectra can be compared directly.
Table 2: Spectral bands of the SeaWiFS and MERIS sensors
| MERIS |
| BN |
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
10 |
11 |
12 |
13 |
14 |
15 |
| BC |
412.5 |
442.5 |
490 |
510 |
560 |
620 |
665 |
681.25 |
705 |
753.75 |
760 |
775 |
865 |
890 |
900 |
| BW |
10 |
10 |
10 |
10 |
10 |
10 |
10 |
7.5 |
10 |
7.5 |
2.5 |
15 |
20 |
10 |
10 |
| SeaWiFS |
| BN |
1 |
2 |
3 |
4 |
5 |
- |
6 |
- |
- |
- |
7 |
- |
8 |
- |
- |
| BC |
412 |
443 |
490 |
510 |
555 |
- |
670 |
- |
- |
- |
765 |
- |
864 |
- |
- |
| BW |
20 |
20 |
20 |
20 |
20 |
- |
20 |
- |
- |
- |
40 |
- |
40 |
- |
- |
Note: BN = Band Number, BC = Band Centre (nm), BW = Band Width (nm)
It can be seen that the MERIS spectra of the eight algal bloom classes are quite distinct from each other. For many algal bloom types, the spectra can be differentiated visually from their shapes around the chlorophyll absorption band at 670 nm. In comparison, the SeaWiFS does not have spectral bands beyond 670 nm. Hence, it is expected that SeaWiFS will fare poorer in terms of accuracy in classification of the algal bloom types.
4. Classification of Algal Blooms from Reflectance
An algorithm based on the singular value decomposition (SVD) technique (Danaher and Omongain 1992) has been developed for the detection and classification of algal bloom types from reflectance data. This algorithm is a type of supervised classification technique. In this algorithm, a "key vector" V
i(
l) for each algal bloom class labelled by the subscript i is first determined from the reflectance spectra of the algal bloom class of interest measured during the field trips. This key vector acts as a sort of template for this class of algal bloom. A given measured spectrum R(
l) to be classified is then "matched" to this key vector using the dot-product operation to give a key value w
i. Mathematically, the dot-product operation is represented by the formula:
Ideally, if the spectrum R(
l) belongs to class-i, then w
i=1, otherwise w
i=0. Using a training set of spectra of known classes, the key vector for each of the nine classes (8 algal bloom classes + 1 reference sea water class) are obtained using the singular value decomposition technique. The key vectors are then matched to each of the unknown spectra R(
l) to be classified, using the dot-product operation. In this way, each spectrum is transformed into a vector of nine key values. The results of supervised minimum distance classification in the key value space are shown in Tables 3 and 4 for the SeaWiFS and MERIS sensors respectively. For comparison, the results of supervised minimum distance classification using normalized spectral values of the respective sensors are shown in Tables 5 and 6.
Figure 1: Simulated SeaWiFS and MERIS spectra for the reference sea water class (class 1) and the 8 algal boom classes (Class 2 to 9)
