Remote sensing of total suspended solids
in penang coastal waters, malaysia
Analysis of TM Data
A sub-scene for each date was extracted covering an image area of 1200 pixels by 1200 pixels for analysis. The images were rectified to the corresponding map (BA Chart no. 1366) to determine their geographical coordinates by using the second order polynomial coordinate transformation. The digital numbers (DN) for bands 1, 2, and 3 at the sample stations were extracted for analysis. These DNs were then converted to radiance unit and effective at satellite reflectance. The simple solar angle correction was performed to the data sets. The atmospheric correction based on the darkest pixel method was employed in the present study (Lathrop et al. 1991, Keiner and Yan (1998), Allee and Johnson 1999).
The relationship between the sea truth reference data (TSS) and the TM bands was first examined using the combined data from both image dates for multi-date analysis. Points situated in the shallow water areas and cloud pixels were removed. The TM signals were then regressed against the suspended solid concentrations using our developed algorithm. Other forms of water quality algorithms were also tested with these data sets and their accuracies were compared with that of the proposed algorithm. For each regression model the correlation coefficient, R, and the root-mean-square deviation, RMS, were noted.
Maps of these water quality parameters were then generated using the coefficients obtained from the regression analysis of our proposed algorithm. Land and cloud areas were masked out using the threshold values of band 4 and thermal band data. The water quality images were geometrically corrected through resampling process using the nearest neighbor method. Image smoothing was performed to each map using the median filter to remove random noise while preserving high frequency features (edges). The generated maps were colour-coded for visual interpretation.
Results and Discussions
Figure 1 shows the plot of the relationship between TM signals versus TSS concentration. As the concentration increases, the response from each TM band also increases. Other investigators using remote sensing data in the visible channels for suspended sediment studies showed similar characteristics (Schiebe et al. 1992, Choubey and Subramaniam 1992). The trend suggests that the non-linear relation is preferred by the data set.
The single band method was found to be less accurate. The calibration results showed that when a single independent variable was used the polynomial form gave better accuracy than the simple linear model. The accuracy was observed to improve when higher order series were used. Generally the accuracy increased when more spectral bands and higher order series were included in the regression analysis. The proposed model used these criteria and therefore produced superior results using different sets of data transformations. Table 1 shows the comparative performance of the algorithms.
For this multi-date analysis, the use of combined raw data did not produce satisfactory results. Solar angle corrected radiance and exoatmospheric reflectance provided some improvement to the results. Figure 2 shows the regression results using the proposed algorithm for TSS using the solar angle corrected radiance data. The results suggest that the atmospheric conditions for the two scenes were nearly the same. Atmospherically corrected data did not appear to improve the accuracy. The most probable cause of the error may be the incorrect choice of the clear deep water radiance for each TM band from the two acquired scenes.
Figure 1 TM radiance versus TSS concentration.
Table 1 Regression results using different forms of algorithms for TSS and chlorophyll
| Algorithm | R | RMS (mg l-1) |
| P=ao+a1B1+a2B12+a3B13 | 0.80 | 54.91 |
| P=ao+a1B3+a2B32+a3B33 | 0.85 | 48.94 |
| P=ao+a1lnB1+a2lnB12+a3lnB13 | 0.75 | 61.09 |
| P=ao+a1lnB3+a2lnB32+a3lnB33 | 0.84 | 49.25 |
| P=ao+a1(B1/B2)+a2(B1/B2)2+a3(B1/B2)3 | 0.78 | 57.05 |
| P=ao+a1(B2/B3)+a2(B2/B3)2+a3(B2/B3)3 | 0.90 | 41.09 |
| P=ao+a1ln(B1/B2)+a2ln(B1/B2)2+a3ln(B1/B2)3 | 0.83 | 53.71 |
| P=ao+a1ln(B2/B3)+a2ln(B2/B3)2+a3ln(B2/B3)3 | 0.90 | 41.77 |
P=a0+a1B1+a2B2+a3B1B2+a4B12+a5B22+
a6B12B2+a7B1B22+a8B12B22 | 0.89 | 44.09 |
P=a0+a1B1+a2B2+a3B3+a4B1B2+a5B1B3+
a6B2B3+a7B12+a8B22+a9B32 (proposed) | 0.95 | 28.17 |
