|
Water quality mapping using multi-date images from digital camera
Regression Algorithm
TSS concentration can be obtained by solving the two simultaneous equations to get the series of terms R1 and R2 that is given as
P = ao+a1R1+a2R2+a3R1R2+ a4R12+ a5R22+ a6R1 2R2+ a7R1 2R22+a8R12R22+… (3)
where aj, j = 0, 1, 2, … are the coefficient for equation (3) that can be solved empirically using multiple regression analysis. This equation can also be extended to the three-band method given as
P = eo+e1R1+e2R2+e3R3+ e4R1R2+ e5R1R3+ e6R2 R3+ e7R12+e8R22+e9R32 ------------------(4)
where the coefficient ej, j = 0, 1, 2, … can also be solve empirically. Data analysis and results The colour digital images of the study area captured by the digital camera contained of 1792 x 1200 pixels. The images were separated into three bands assigned as red, green, and blue bands. The separated bands were stored in raw data format in order to facilitate further analysis using in-house programs and PCI software. Ground control points (GCP) were determined by using a geometrically corrected SPOT satellite map as the reference geocoded image. Figure 2 below shows the image of 9 March 2002 used in this study. 
The image of 28 October 2001 was taken obliquely. The view angle correction was first performed to correct the angular dependence of image brightness. Then the multi-date correction technique was carried out. The vertical image captured on 9 March 2002 (Figure 2) was used as the reference image. Different types of materials visible in the reference image and the image to be corrected (image on 28 October 2001) were selected as correction targets. We assumed the reflectance of these targets did not change with time. The DNs of the targets in the second image were related to the reference image. The relationship was then used to adjust the DN values of the second image. This correction forced the image to have the same atmospheric conditions and the multi-date analyses were then performed using these data sets. The DN values corresponding to the water sample locations were extracted from both images. The DNs for window size of 7 by 7 were used for the calibration of the algorithm. This window size was selected because the data set produced higher correlation coefficient and lower RMS value. Half of the data were chosen randomly for the calibration data set. The remaining data were used in the validation analysis. The relationship between TSS and DN for the calibration data set is shown in Figure 3. Superior result was produced by the proposed model, which achieved the correlation coefficient of about 0.9988 as shown in Figure 4. The Accuracy of the proposed calibrated algorithm was then assessed using the validation data set. The analysis produced high correlation coefficient (R=0.95) between the predicted and the sea-truth values as shown in Figure 5. 
Various water quality algorithms were also tested and their accuracies (R and RMS values) were noted (Keiner and Yan 1998, Pulliainen et al. 2001). Table 1 shows comparison for various algorithms. The proposed algorithm produces higher correlation coefficient (R=0.9988) between the predicted and the measured TSS values and lower RMS value (1.5233 mg/l) compared to other algorithms. The TSS map was generated using the proposed calibrated algorithm. The map was then geometrically corrected by using the cubic convolution method. This method was used because it produced a smoother geocoded image. The generated map was then filtered by using 5 by 5 pixels average for removing the random noise. Then, the colour-code was produced for visual interpretation as shown in Figure 6. 
Table 1. Regression results using different forms of algorithms for TSS
Note: B1, B2 and B3 are the digital numbers for red band, green band and blue band respectively The contour map was plotted using sea-truth data with Kriging method. Quantitative comparison between contour map plotted using the sea-truth data and image generated by the algorithm gives acceptable result. Both images show almost the same pattern as can be seen in Figure 6 and 7. This shows the capability of the algorithm to generate TSS concentration map. Generally, TSS is higher in the river estuary closer to the beach as shown by both maps. However, the map generated using the algorithm shows that the concentration of TSS near to the beach is more than 250mg/l, whereas the contour map plotted based on sea-truth data gives the range of 201 to 250 mg/l. The difference is due to the limited sea-truth data used to generate the contour map by the interpolation technique. |