Using Spectral and Spatial Information for Oil Spill Detection in Multi-spectral Imagery
4. Experimental Result
4.2. Real Image Scene
(a) Scenario 1 (b) Scenario 2
(c) Scenario 3 (d) Scenario 4
Fig.11 Threshold results for Case 2
Table 3
Performance analysis for simulation case 2 Case Correct Detection Errorpixels % pixels % 1 Before 411 72.36 347 0.87
After 342 60.21 226 0.57 2 Before 372 65.49 446 1.16 After 318 55.99 250 0.65 3 Before 429 75.53 294 0.77 After 353 62.15 215 0.56 4 Before 420 73.94 294 0.77 After 408 71.83 160 0.42
4.2. Real Image Scene
For the real image scene, we first remove the land area as a preprocess step. Based on the strong absorption of the water in IR region, we separate the land area from the sea by a threshold. Figure 12 shows the original images with land area removed.
For the first scenario, the RX algorithm is applied on these three band multispectral images with only the spectral information according to Eq. (1). Figure 13(a) shows the RX result, the oil
slick gives higher score and shows as the bright area in the image. After applying a threshold, as shown in Fig. 13(b), oil slick is detected as white pixels along with other interferers. After the morphologic operations “closing and opening”
to remove the interference, it still shows false alarm region in the middle and bottom of the image (Fig. 13(c)).
The second scenario is to use only the spatial feature information for oil slick detection.
We first calculate the 33 spatial feature images for the three original images (Fig. 14 and 15), and select 18 images for the experiment. They are Mean and Variance of Gray Level Histogram;
Coarseness and Variance of Texture Feature Coding Method for all three bands.
(a) (b) (c)
Fig.12
Original SPOT-1 images with land area removed(a) (b) (c)
Fig.13 Scenario 1
(a)Mean (b)Variance (c)Kurtosis
(d)Mean (e)Variance (f)Kurtosis
(g)Mean (h)Variance (i)Kurtosis
Fig.14 Gray Level Histogram: (a)-(c) from band;
(d)-(f) from band2; (g)-(h) from band3
Coarseness0 Coarseness1 Homogeneity0 Homogeneity1
Mean Convergence0
Mean
Convergence1 Variance0 Variance1
Coarseness0 Coarseness1 Homogeneity0 Homogeneity1
Mean
Convergence0 Mean
Convergence1 Variance0 Variance1
Coarseness0 Coarseness1 Homogeneity0 Homogeneity1
Mean Convergence0
Mean
Convergence1 Variance0 Variance1
Fig.15 Texture Feature Coding Method
(a) (b) (c)
Fig.16 Scenario 2
(a)Mean (b)Variance (c) Kurtosis
Fig.17 Gray Level Histogram form RX result.
(a)Coarseness0 (b)Coarseness1 (c)Homogeneity0 (d)Homogeneity1
(e)MC0 (f)MC1 (g)Variance0 (h)Variance1
Fig.18
Texture Feature Coding Method(a) (b) (c)
Fig.19 Scenario 3
(a) (b) (c)
Fig.20 Scenario 4
Fig.21
Original SPOT-1 images(a) Scenario 1 (b) Scenario 2
(c) Scenario 3 (d) Scenario 4
Fig.22 Detection results comparison
We applied RX algorithm and morphologic operation as the first scenario on these 18 images as shown in Fig. 16. Compare to Fig. 13, it is clearly shown that the threshold image, Fig.
16(b), has successfully reduce the interferences while maintaining the detection performance.
Therefore, the spatial features did provide benefits in separating the interference from the oil slick.
Since there are some useful information in spatial features. In the third scenario, we extract spatial informtion only from the RX result in Fig.
13(a). The six spatial features are Mean and Variance of Gray Level Histogram; Coarseness;
and Variance of Texture Feature Coding Method.
Combining with the original three spectral images, we have a 9-band image cube.
We applied RX algorithm and morphologic operation as the first scenario on these images as shown in Fig. 19. Because the spatial features are extracted from the RX result which uses only the original three bands, it still has some interference mis-detected in the lower half image in Fig. 19(b). But the performance is significantly improved in the upper half image comparing with Fig. 13(b).
For the last scenario, we combine the spatial textures extracted directly from the original images (scenario 2) and the original 3 bands. We applied same procedure on this 21-image cube, and the result is shown in Fig.
20. It is noticed that this combination has lowest false alarm probability among all four scenarios.
We further overlap the boundary of detected oil slick area to the original image in Figs. 21 and 22. The result of scenario 1 with only the spectral information detected interference as false alarm areas in Fig 22(a), while the other three scenarios with spatial features detect only the oil slick area after morphologic operation. The Fig. 22(c) is the result of scenario 3 with original 3-band images and the spatial features extracted from the scenario 1, and the lower-left part of the ship is detected as oil slick. Both scenarios 2 and 4 adopt spatial feature images extracted from original 3-band images and they show similar results in Fig. 22(b) and (d). But comparing with Fig. 16(b) and 19(b), scenario 4 with both spectral and spatial information performs the best.
5. Conclusion
RX algorithm with only spectral information can detect the oil spill on sea surface, but still reveals some interference produced by marine phenomena. In this study, we demonstrate that the combination of Spatial Feature Information with spectral information, the performance RX algorithm can be improved.
Finally, a mathematical morphology method is applied to the image which further filters out the interference of the sea phenomena. The performance analysis with synthetic scene simulation shows the detection with both spectral and spatial features outperforms the scenario with only spectral or spatial information is available. The experiment with real image scene also supports the finding.
Therefore, the proposed procedure with both spectral and spatial information is indeed a good method for oil slick detection.
Reference
Brown, C.E., Fingas, M.F., Goodman, R.H., Mullin, J.V., Choquet, M. and Monchalin, 2000. Airborne Oil Slick Thickness Measurement, in Proceedings of the Fifth International Conference on Remote Sensing for Marine and Coastal Environments, Environmental Research Institute of Michigan, Ann Arbor, Michigan, pp. I219-224.
Chen, H.T. (陳獻廷)，2006。利用高光譜影像作異 常物偵測，國立中央大學太空科學研究所碩士 論文。
Christoyianni, I., Dermatas, E. and Kokkinakis, G., 2000. Fast detection of masses in computer-aided mannography, IEEE Signal Processing Magazine, pp. 54-64.
Del Frate, F., Petrocchi, A., Lichtenegger, J. and Calabresi, G., 2000. Neural Networks for Oil Spill Detection Using ERS-SAR Data.
IEEE Transactions on Geosciences and Remote Sensing, Vol. 38, No. 5, pp.
2282-2287.
Fingas, M.F. and Brown, C.E., 2000. Review of Oil Spill Remote Sensing, in Proceedings of SPILLCON, Australian Marine Safety Authority, Sydney, Australia.
Huang, C.C. (黃俊忠)，2006。利用 X 光乳房攝影 產生之紋理特徵影像在腫瘤偵測上之研究，國 立中央大學資訊工程研究所碩士論文。
Kelly, E.J., 1989. Adaptive detection and
parameter estimation for multidimensional signal models, MIT Lincoln Laboratory, MA, Tech. Rep., Apr., pp. 848.
Liao, Y.C. (廖友千)，2001。紋路特徵值分析應用 於乳房 X 光片攝影之腫瘤偵測，國立成功大 學資訊工程學系碩士論文。
Liu, C.H. (劉建華)，2007。台灣南部海域溢油動態 資料庫-應用於海洋污染事故應變模擬分析，
國立中央大學環境工程研究所碩士論文。
Lopez, L., Moctezuma, M. and Parmiggiani, F., 2005. Oil spill detection using GLCM and MRF, IEEE, pp. 1781-1784.
Reed, I. and Yu, X., 1990. Adaptive Multiple-Band CFAR Detection of an Optical Pattern with Unknown Spectral Distribution. IEEE transactions on acoustics.
speech. and signal processing, Vol. 38, No.
4, pp. 1760-1770.
Taylor, S, 1992. 0.45 to 1.1 μm Spectra of Prudhoe Crude Oil and of Beach Materials in Prince William Sound, Alaska, CRREL Special Report No. 92-5, Cold Regions Research and Engineering Laboratory, Hanover, New Hampshire, 14 p.
1國立台灣海洋大學應用地質研究所博士候選人 收到日期:民國 98 年 01 月 01 日
2國立台灣海洋大學應用地質研究所教授 修改日期:民國 98 年 02 月 17 日
3國立中央大學太空科學研究所碩士生 接受日期:民國 98 年 03 月 09 日
4國立中央大學太空及遙測研究中心副教授