MULTI-TEMPORAL SATELLITE IMAGES WITH BATHYMETRY CORRECTION FOR MAPPING AND ASSESSING SEAGRASS BED CHANGES IN DONGSHA ATOLL

Transcription

1 MULTI-TEMPORAL SATELLITE IMAGES WITH BATHYMETRY CORRECTION FOR MAPPING AND ASSESSING SEAGRASS BED CHANGES IN DONGSHA ATOLL Chih -Yuan Lin and Hsuan Ren Center for Space and Remote Sensing Research, National Central University, Taiwan Keywords: Dongsha Atoll, Multispectral Images, Bathymetry Correction, Seagrass Classification ABSTRACT: The Dongsha Atoll is the first ocean national park in Taiwan. It is a 400 kilometer-square atoll with coral reef ecosystems and high biodiversity. Because it has only one small island in the western of atoll and has not opened to public, there are very few human activities and nature is well preserved. To survey this large area is not an easy task, and remote sensing technique provides an efficient and economic approach to survey this area. The multispectral satellite images are useful for assessing sea bottom materials. With multi-temporal images, the changes can also be tracked. A huge event has been observed in summer 2014, more than 25 kilometer-square of seagrass in northern atoll disappear within three months, and the habitat changes from seagrass to sand and coral reef debris. In this study, satellite images with finer temporal resolution will be analyzed to track the speed of changes. However to classify bottom material directly from spectral information has some difficulties, because spectral information is also mixed with other factors, including water absorption coefficient and water depth. In this research, sun glint and water depth correction will be conducted to remove the effects from atmosphere and bathymetry. And image classification for seagrass is followed with multi-source satellite images to track the disappearance and recovery of the seagrass area. The satellite images for experiments include FORMOSAT-2, SPOT-6 and LandSat Introduction Dongsha atoll national park is located in the South China Sea. It became the seventh national park in 2007 and also the first ocean national park in Taiwan. With 25 kilometers in diameter, the atoll covers over 400 kilometer square area, and it is rich of coral reefs and marine resources. Because this national park is not yet open to public, it has very few human activities there. Start from 2012, our group joined the research team for Marine National Park Headquarters, combine remote sensing images with in situ survey for sea bottom material classification in Dongsha atoll. In the 2014 summer, we observed a huge event for seagrass in DongSha atoll, more than 25 kilometer square of seagrass habitat became sand and coral reef debris, which has never be recorded for the past 20 years. Figure 1. The DongSar Atoll and multi- temporal imagery. 1.1 Objective

2 The benthic species generally are sensitive to the nature hazard, but there was neither typhoon nor earthquake near Dongsha between July and September, This is not seasonally change because it is a single event for the past 20 years. Before establishing the hypothesis why the seagrass disappear, we need to know the speed of its disappearance, and monitoring the changes with multi-temporal remote sensing images can provide some hints. Therefore, the main goal of this research is to collect as many remote sensing images as possible from different sensors between July and September, 2014 and mapping the seagrass area with high classification accuracy. 1.2 Material We have collected 9 images between March 2014 and May 2015 from Formosat-2, SPOT-6 and Landsat-8, as shown in Fig. 2. The images from these three sensors have different resolutions and frequency bands, but they all have visible bands (R, G, B) and near infrared band. Since infrared light will be absorb by 20 cm water, it is used to correct sun glint. The visible bands are used to map the bottom materials. Figure 3 shows the enhanced images. Figure 2. The satellite imagery times. Figure 3. Multi temporal imagery The DongSar Atoll and multi- temporal image. (A) In-situ depth data. (B) FS /03/13. (C)Landsat /06/10. (D)Landsat /08/30. (E) FS /11/02. (F) Spot /2/18. (G) Spot /5/28. (H) FS /05/ Methodology

3 Our proposed method contains three steps: sun glint correction, bathymetry calibration and bottom material classification. The details are discussed in following sections. 2.1 Sun glint correction The coastal environment study often has sun glint effect caused by the wave on water surfaces. It is a serious confounding factor for radiance and reflectance. Generally, infrared band can be used to remove the sun glint effect. We adopted the method of Hedley (2005), to correct sun glint effect. (1) 2.2 Bathymetry calibration Bathymetry calibration is a common technique in the shallow costal area. The water depth can be estimated from the remote sensing data by the exponential relationship of the difference between sea bottom radiance and deep water radiance. Lyzenga equation is the fundamental formula of inversion the water depth. The inversion result is influent by the water attenuation coefficient and substrate type. R R R e R (2) 2 Zg s ( 0 ) where R is the radiance return to sensor, R 0 is the bottom radiance, R is the deep water radiance, g s is water attenuate coefficient and Z is water depth. With bathymetry information, we could calibrate the water depth and derive bottom radiance. 2 Zg s R0 ( R R ) e R (3) 2.3 Maximum likelihood classification After bathymetry calibration, bottom radiance is derived. For assessment the seagrass change and other bottom material, maximum likelihood classification is adopted to classify bottom materials. Each material is modeled by a Gaussian distribution and each sample is classified to the class with highest probability. (4) (5),

4 3. RESULTS AND DISCUSSION Compare the images from Formosat-2 in 2014/03/13 and 2015/5/28 Formosat-2, seagrass disappearance can be easily found. To know the time and speed of the disappearance, a series of images from different sensors are collected. In our experiments, the sun glints are first corrected for those images as shown in Fig. 4, followed by bathymetry calibration in Fig. 5. The classification results by Maximum Likelihood Classifier are shown in Fig.6, and we can observe the seagrass disappearance starts in June 2014 in north-east of atoll, and it continues in northern part before August, and have not recovered since then. Figure 4. The detail scene in the image before sun light correction (Left), after correction (Right). (Top) Reduce the sea surface. (Mid) Raise the sand and seabed contrast. (Bottom) Seagrass and other material r become more clear

5 Figure 5. Use water depth retrieval the bottom spectral from multi-temporal satellite images. (A) FS /03/13 derive depth.(b) Landsat /06/10 derive depth. (C) Landsat /08/30 derive depth. (D) FS /11/02 derive depth.(e) Spot /2/18 derive depth. (F) FS /05/28 derive depth. (G) Spot /05/28 derive depth. Figure 6. (A) Formosat /0313 classification. (B)Landsat /0610 classification. (C)Landsat /08/30 classification. (D) Formosat /11/02 classification. (E)Spot /02/18 classification. (F)Spot /05/28 classification.(g)formosat /05/28 classification. For the statistics assessment, we calculate the area for each class from the classification result as in Table 1. The area of seagrass decreased tremendously between June and August Due to the wave and tidal changes, the boundary

6 of the atoll rock is difficult to be detected. The average of overall accuracy is 77.39% and average Kappa Coefficient is There are still some error and misclassification, which might be caused by the tidal environment usually occur near coastal line. Table 1. Classification Result. Maximum likelihood classification area Satellite Formosat-2 Landsat-8 Landsat-8 Formosat-2 time 2014/03/ /6/ /08/ /11/02 Sea (61.12%) (61.7%) (62%) (63.3%) Seagrass (11.45%) (11.2%) (7%) (4.7%) Mud & Sand sediment Mixture (5.73%) (8.4%) (9.23%) (11.93%) Coral habitat (17.34%) (11.6%) (11.6%) (15.6%) Coral & Rock (1.24%) (2.2%) (1.2%) (5.3%) Other Mixture Sediment (1.8%) (4%) (5) (2.83%) Cloudy none (1.24%) (0.7%) (4.4%) note none none Cloudy day none Unit: square kilometer (area / sum of study area ) Satellite Spot-6 Spot-7 Formosat-2 time 2015/02/ /5/ /05/28 Sea (63.33%) Seagrass 9,00 (4.79%) (61.35%) (4.56%) (62%) (5%) Mud & Sand sediment Mixture (14.48 %) (12.18 %) (8.7%) Coral habitat (14.79%) Coral & Rock 2,581 (1.3%) Other Mixture Sediment (1.18%) (15.6%) 5.18 (2.76%) 7.06 (3.76%) (14.88%) (3.86%) (4.1%)

7 Cloudy none none none note none none none Unit: square kilometer (area / sum of study area ) Table 2. Classification Accuracy All classification accuracy assessment TIME Spot-6 Spot-7 FS /02/ /05/ /05/28 Overall Accuracy 80.04% % 80.11% Kappa Coefficient TIME FS-2 Landsat-8 Landsat-8 FS /03/ /6./ /08/ /11/02 Overall Accuracy % 78.71% % 73.91% Kappa Coefficient Total Area: Km^2 Average Over-all Accuracy: Average Kappa Coefficient: : From the Formosat-2 image in 2014/03/13 to Landsat-8 images in June and August, a large area (about 25 Km^2) of seagrass disappeared within 3 months. Although the resolution of Landsat-8 is 30 meter, relative low compare to Formosat-2 and SPOT-6/7, it can clearly recognize this event. From Figure 6, images after November 2014 indicate that the seagrass area is still not recovered. 4. Conclusion and Future Work In this study, our proposed method can provide good results for seagrass classification and track the changes. It can remove sun glint and calibrate bathymetry to reduce the solar and water depth effects, so the supervised maximum likelihood classifier can extract the seagrass. Although images from Landsat-8 Formosat-2 and Spot-6/7 have different spatial resolutions and spectral frequencies, they are able to detect the change in larger area. Compare with the long term change result we estimate the seagrass habitat lose about 19%. In the future, we would like to include tidal level and atmospheric calibration in further reduce error. We will also develop algorithm for transformation between SPOT-7 and Formosat-2, since they both collect images on May 28, We expect to improve the consistency between the classification results between different sensors.

8 5. Reference Pu, R., Bell, S., & Meyer, C. (2014). Mapping and assessing seagrass bed changes in Central Florida's west coast using multi-temporal Landsat TM imagery. Estuarine, Coastal and Shelf Science, 149, Hedley, J. D., Harborne, A. R., & Mumby, P. J. (2005). Technical note: Simple and robust removal of sun glint for mapping shallow water benthos. International Journal of Remote Sensing, 26(10), Lyons, M., Phinn, S., & Roelfsema, C. (2011). Integrating Quickbird multi-spectral satellite and field data: Mapping bathymetry, seagrass cover, seagrass species and change in Moreton Bay, Australia in 2004 and 2007.Remote Sensing, 3(1), Lyzenga, D. R. (1978). Passive remote sensing techniques for mapping water depth and bottom features. Applied optics, 17(3),