Evaluation of Underwater Spectral Data for Colour Correction Applications

Transcription

1 Proceedings of the 5th WSEAS Int. Conf. on CIRCUITS, SYSTEMS, ELECTRONICS, CONTROL & SIGNAL PROCESSING, Dallas, USA, November 1-3, Evaluation of Underwater Spectral Data for Colour Correction Applications JULIA ÅHLÉN University of Gävle Dept. of Mathematics, Natural and Computer Sciences Gävle SWEDEN EWERT BENGTSSON Uppsala University Centre for Image Analysis Lägerhyddsv. 3, SE Uppsala SWEDEN DAVID SUNDGREN Royal Institute of Technology Dept. of Computer and Systems Sciences Forum 100, Kista SWEDEN Abstract: The inherent properties of water column usually affect underwater imagery by suppressing high-energy wavelengths. One of the inherent properties, diffuse attenuation, can be estimated from multi or hyper spectral data and thus give information on how fast light of different wavelengths decreases with increasing depth. Based on exact depth measurements and data from a spectrometer incoming light on an object can be calculated and diffuse attenuation coefficient can be estimated. In this work the authors introduce a mathematical model that suggests the most stable wavelengths, which corresponds to estimated coefficients, based on spectral information from each depth. These values are then used in reconstruction of colours in underwater imagery. Since there are no digital hyper spectral cameras yet we are for the time being confined to point data, but the method is general and we show how it can be applied on multi spectral images. Key Words: Water, Colour Correction, Spectral Data 1 Introduction Coral reefs and related ecosystems are threatened as a result of developing economies and increasing coastal populations. As much as 58% of the world s coral reefs are at risk due to human activity such as coastal development, destructive fishing, overexploitation, marine pollution, runoff from deforestation and toxic discharge from industrial and agricultural chemicals [1]. Methods for careful monitoring, planning and management becomes essential [6]. Damaged corals are bleached, so we can use underwater imagery with correct colour representation to establish the health of corals. However, colours are distorted under water, so we need a method for colour restoration. 1.1 Techniques for monitoring coral reefs Remote surveillance from satellites such as Landsat and SPOT aid a global estimate on coral health if underwater features can be properly identified. Hyperspectral satellite-borne sensors allows large geographic areas to be covered with high spectral resolution at potentially lower cost than higher spectral resolution airborne sensors. However in situ measurements are required when discrimination of small featured benthic habitats are needed. Since the spatial resolution of satellite imagery is not satisfactory - a user might have to accept 10 m spatial resolution in exchange for 100 narrow and contiguous spectral bands - it is important to focus on fundamental issues such as spectral distinction between optically similar substrates using in situ data rather than using remotely sensed imagery immediately. Initially, it is difficult to perform an assessment of the accuracy of a classification of a remotely sensed image due to the large geographic area covered and the difficulties involved with geopositional accuracy. Mixing of reflectance signatures within each pixel such that it is difficult to determine the pure reflectance characteristics of only healthy coral or only sea grass is another problem. If high spatial resolution imagery is captured using an airborne sensor, then the pixel size could be smaller than 1 m, but there can still be several substrate types present within this small area in a typical coral reef environment. Finally, remote sensing is espoused as an ideal tool for resource management and ecosystem monitoring, but the fundamental research is not complete in the areas of water column correction and substrate identification, so its capabilities may have been oversold thus disappointing potential users [3]. For the past 13 years, Reef Relief s founder, Craig

2 Proceedings of the 5th WSEAS Int. Conf. on CIRCUITS, SYSTEMS, ELECTRONICS, CONTROL & SIGNAL PROCESSING, Dallas, USA, November 1-3, Quirolo, has documented changes in the coral communities of the Key West area with a non-intrusive photographic survey. A series of pictures taken over months and years monitor growth and overall health [8]. An extensive usage of digital camera for monitoring corals provides high spatial resolution imagery and provides an economical way of monitoring coral reefs. There are several drawbacks of this data collection method such as corrupted colours due to absorbtion of light by water column, high demands on divers photographing skills, and difficulties to maintain camera under the water [7]. The main demand on a digital camera for scientific purposes is correct colour representation, however the accurate colorimetric reproduction is often not the primary goal for camera manufacturers [12]. Instead, they strive to make the cameras create images that are visually pleasing by using a rendering function, which makes grass greener and skies bluer [9]. In [14] a method for estimating the effect of these functions is presented. 1.2 Water column effects on underwater imagery The water column significantly affects the sensed signal through wavelength-specific augmentation and attenuation, which does not always follow Beer s Law [2] of logarithmic extinction of energy. Radiative transfer models sensitive to varying substrate brightness, water depth, and water quality are needed to account for the variable effects of the water column [4]. Based on intensity values of the image we can estimate the downwelling irradiance and diffuse attenuation coefficients K d of the diving site. Strictly speaking, K d is not a property of the water itself but rather a descriptor of the underwater light field that varies with depth, solar altitude, and time. The authors of [15] are arguing that the absorption coefficients can give an indication on what K d values should not be used in colour reconstruction of the images, but no discussion is given on which wavelengths that should be used to compute the K d values needed in reconstruction of colours. Addressing this issue we present a Stability Model which will give a value range for wavelengths used to compute K d values that are as stable as possible in terms of variation with increasing depth. Although the main raison d être for the method is examining the health of corals, it is generally applicable to any type of water. Below we show results based on measurements collected in waters off Sweden and the result of colour correction of underwater images from Portugal. 2 Data Collection and Instruments Data was collected off the coast of Gävle, Sweden, in May Solar altitude was 42. For this study we used a spectrometer [11] and took measurements of spectra of a gray reflectance plate [10] under the water. The gray plate was attached to device that held the plate stable and at a right angle to the surface. The integration time was the same for each measurement. We measured at depths between 0 m and 4 m with 20 cm intervals. The spectrometer registered pointwise the intensity counts in 1024 spectral channels from nm to nm with the wavelength resolution of 0.29 nm. At the site water conditions are such that it is almost opaque close to the surface. This is due to a high concentration of organic matters and dissolved particles. 3 Establishing K d values Light that penetrates the ocean surface undergoes severe energy loss and changes direction due to small particles dissolved in the water. We are estimating the rate of extinction of light of particular wavelengths when the depth increases. Here we use intensity counts from a spectrometer, but it is also possible to use intensity values from a digital image to produce poor man s K d values for the three channels red, green and blue [13]. Based on the values of the intensity counts from each measured depth we obtain K d values as a function of wavelength λ and the difference zz 1 between two depths z and z 1, see Equation 1. Equation 1 is an approximation of Beer s law that is applicable when differences in depth are small. K d (zz 1, λ) = I(λ, z 1) I(λ, z) zz 1 I(λ, z), (1) where I(λ, z) is the intensity of wavelength λ at depth z. 3.1 Development of Stability Model The curves of K d as a function of wavelength are sometimes jagged, due to particles in the water, reflection of solar light on waves, measurement errors and other factors, see Section 4 and Figures 1, 2, 3 and 4. To minimise the effect of the arbitrariness when choosing one wavelength per spectral band we seek the subinterval where the curve is as smooth as possible. When determining the most stable and smooth wavelength interval for each spectral channel we want to minimise the total rate of change in small subintervals, i.e. in effect find the least jagged interval, see Equation (2).

3 Proceedings of the 5th WSEAS Int. Conf. on CIRCUITS, SYSTEMS, ELECTRONICS, CONTROL & SIGNAL PROCESSING, Dallas, USA, November 1-3, Figure 1: Curves of the intensity counts as a function of wavelength for the depths from 0.2 to 2 m. Figure 4: Curves of the K d values as a function of wavelength for the depths from 2.2 to 4 m. We divide the whole visible spectra in a red, green and blue intervals according to the following: red: nm, green: nm and blue: nm, [5]. Further we divide each of these intervals in small subintervals R i, G i and B i λ 1, λ 2,..., λ n, where n = 22 for R i, n = 21 for R g and n = 15 for R b. For each of red, green and blue, we seek the subinterval R i, G i and B i where n 1 depth i=1 K d (λ i ) K d (λ i+1 ) (2) Figure 2: Curves of the intensity counts as a function of wavelength for the depths from 2.2 to 4 m. is minimal. The subintervals, in turn, in our case consists of ten different wavelengths 0.29 nm apart. For the subinterval where the expression in Equation 2 is minimal, we can use any of the ten wavelengths since they all correspond to roughly the same K d value. This K d value is then set into colour correction algorithm, where Beer s Law is involved and the image can be restored in terms of colours. Figure 3: Curves of the K d values as a function of wavelength for the depths from 0.2 to 2 m. 4 Results and Discussion The developed stability model can be applied to different sets of underwater data, regardless of e.g. location and water type. The jagged reflected radiance spectra are to be expected because the solar spectrum is distorted due to Fraunhofer lines and atmospheric absorption lines. The K d should be highest at 400 and 800 nm and have a minimum around 500 to 600 nm. However, these K d values are affected by high concentration of non-organic dissolved particles that are

4 Proceedings of the 5th WSEAS Int. Conf. on CIRCUITS, SYSTEMS, ELECTRONICS, CONTROL & SIGNAL PROCESSING, Dallas, USA, November 1-3, Figure 5: Original image taken at 6 m of depth and pre-processed to eliminate the build in camera functions. Figure 7: Corrected image with the arbitrarily chosen Kd -s. Channel Red Green Blue Image corrected with stability models Kd :s Image corrected with arbitrary Kd :s Table 1: The absolute value of the difference with the reference image and the colour corrected images Figure 6: Corrected image with the Kd -s obtained with the stability model. commonly present in waters close to the east coast of Sweden. Below we show the effect of colour reconstruction based on the stability model on imagery from Portugal, where both spectra and digital images are obtained for different depths. The spectra is used to calculate the Kd values which are then set into the approximation of Beer s law, Equation 1, in order to lift up the image to the desired depth by diminishing the effects of severe light absorbtion. In Figure 5 we see an image taken at 6 m depth, which is first preprocessed in order to eliminate the beautifying function built into the digital camera, and in Figure 6 the result of colour correcting under the stability model by using Kd values between 1.8 and 6 m. This means that we see the colours as if the image was taken at 1.8 m depth. In Figure 7 we see the image from Figure 5 colour corrected with arbitrarily chosen Kd values from the three respective channels, as opposed to using the stability model. Since it is difficult to visually decide from these images which Kd s gave the best result, we make a comparison with a reference image taken with digital camera at 1.8 m depth. The gray 99% reflectance plate is present on all images and we compare the average intensity of the gray plate on the image from 1.8 m depth, for red, green and blue channels with that of the two colour corrected images, where the stability model was used and for the image where arbitrary Kd values were used. See Table 1, where the absolute values of the difference between the reference image and each of the colour corrected images are shown; one with the stability model s Kd values and one with the arbitrarily chosen Kd values. Since the absolute values of the differences are smaller where the stability model was used, we claim that Kd values extracted from the stability model should be used for colour reconstruction of underwater images. 5 Conclusion and Summary We have developed a stability model for obtaining suitable Kd values for colour reconstruction purposes. The colour reconstruction method for underwater images gives a simple and economically defendable way for monitoring coral reefs and other marine habitats.

5 Proceedings of the 5th WSEAS Int. Conf. on CIRCUITS, SYSTEMS, ELECTRONICS, CONTROL & SIGNAL PROCESSING, Dallas, USA, November 1-3, Taking images under the water for scientific measurements differs from recreational underwater photography. With our method marine scientists will not have to prepare the equipment necessary to create favourable light conditions. K d values that indicate presence of organic or non-organic matters can be used to extract and precisely show in what depth these particles are situated. For the application of colour correction we are not interested in recognition and classification of such matters, we only need the K d s to proceed with the stability model. An analysis of the obtained K d values suggests wavelength intervals where K d as a function of wavelength is optimally smooth, making these intervals suited for colour reconstruction purposes. For the time being there are no digital hyperspectral cameras which would allow an extensive amount of K d values to choose from, but we argue that the method is applicable on cameras with only three spectral channels. If spectral sensitivity curves are known it is easy to estimate the three K d values representing wavelength range for red, green and blue channels. For future work we would suggest an interpolation of spectral sensitivity curves built into the digital camera with the reflectance profile of the photographed object. This would give a much wider range of K d values to be put into stability model. If several spectral channels are available, our stability model can be applied to these as well. Acknowledgments The authors are grateful to Knowledge Foundation that sponsored this work. We also would like to give special thanks to diving instructor Ricardo Calado that helped the authors to collect the data and associate professor Tommy Lindell for his advises. References: [1] Bryant, D., Burke,L., McManus, J., Spalding,M., Reefs at Risk: a map-based indication of threats to the world s coral reefs. World Resources Institute, Washington, DC, pp. [2] Calculations Using Beer s Law, November, url: law.htm. [3] Green, E., Clark,C., Mumby, P., Edwards, A., Ellis, A., Remote sensing techniques for mangrove mapping. International Journal of Remote Sensing. 19 (5), [4] Holden, H., LeDrew, E., The effects of the water column on hyperspectral reflectance of submerged coral reef features. Bulletin of Marine Science. LeDrew, E., Wulder, M., and H. Holden. Change detection of satellite imagery for reconnaissance of stressed tropical corals. Proc. International Geophysical and Remote Sensing Symposium, Hawaii, USA, July [5] Jones, E., Childers, R., Contemporary College Physics, 2nd Edition, Addison-Wesley Pub Co, Boston, 153 p. [6] Knight, D., LeDrew E., Holden, H., Mapping submerged corals in Fiji from remote sensing and in situ measurements: applications for integrated coastal zone management. Oceans and Coastal Management. 34 (2), [7] Kohler, A., Kohler, D., Underwater Photography Handbook, New Holland Publishers, UK London. [8] Reef Relief a non-profit grassroots membership organization dedicated to preserve and protect living coral reef ecosystems through local, regional and international efforts, url: August [9] Spaulding, K.E., Woolge, G.J., Giorgianni, E.J., Optimizied extended gamut color encodings for scene-referred and output-referred image states. Journal Imaging Science Technology 45 (5) [10] Spectralon, Reflectance Material for Component Fabrication, January, Labsphere, url: [11] Ocean optics, August, url: [12] Vrhel, M., Saber, E., Trussell, H. J., Color image generation and display technologies. IEEE Signal Processing Magazine 22 (1), [13] Åhlén, J., Bengtsson, E., Lindell, T., Color Correction of Underwater Images Based on Estimation of Diffuse Attenuation Coefficients, Proc. PICS Conference, An International Technical Conference on The Science and Systems of Digital Photography including the Fifth International Symposium on Multispectral Color Science, Rochester, NY, May [14] Åhlén, J., Improvement of a Color Correction Algorithm for Underwater Images Through Compensating for Digital Camera Behaviour Proc. Swedish Symposium on Image Analysis, Uppsala, Sweden, March, 2004, pp

6 Proceedings of the 5th WSEAS Int. Conf. on CIRCUITS, SYSTEMS, ELECTRONICS, CONTROL & SIGNAL PROCESSING, Dallas, USA, November 1-3, [15] Åhlén, J., Sundgren, D., Lindell, T., Bengtsson, E., Dissolved Organic Matters Impact on Colour Reconstruction in Underwater Images. Proc. of the SCIA 14-th Scandinavian Conference on Image Analysis, Joensuu, Finland, June, 2005.