Automatic water surface extraction algorithm for Landsat7-ETM+ satellite image over coastal and inland waters

ISSN 2354-0575 Journal of Science and Technology74 Khoa học & Công nghệ - Số 27/Tháng 9 - 2020 AUTOMATIC WATER SURFACE EXTRACTION ALGORITHM FOR LANDSAT7-ETM+ SATELLITE IMAGE OVER COASTAL AND INLAND WATERS Dinh Ngoc Dat1,*, Pham Van Bach Ngoc1, Huynh Xuan Quang1, Mai Thi Hong Nguyen1, Doan Minh Chung1, Bui Trung Thanh2 1 Institute of Space Technology, Vietnam Academy of Science and Technology, Cau Giay, Hanoi 2 Hung Yen University of Technology and Education, Dan Tien, Khoai Chau, Hung Yen * Corresponding author: dndat@sti.vast.vn Received: 15/05/2020 Revised: 20/08/2020 Accepted for publication: 25/09/2020 Abstract: Identifying pixels on the water surface natural water is an important step before applying a specific algorithm to calculate water environment parameters from satellite imagery. A mistake can occur when detecting clouds on turbid water (due to glare) and detecting their shadows on any surface water. Some algorithms exist but their performance is unsatisfactory, especially on turbid waters where cloudless pixels are sometimes classified as clouds or earth, resulting in data loss. This is especially important for the satellite image sensing have high resolution, such as the observations made by the sensor ETM + on Landsat-7, OLI on Landsat-8 or MSI on Sentinel- 2. In this study, we develop a two-step algorithm to extract water pixels (called LS7WiPE) for the Landsat-7 / ETM + sensor based on our experience from the WiPE algorithm for OLI and MSI sensors (Dat et al., 2019). Keywords: WiPE algorithm, Landsat-7/ETM+ satellite imagery, phân tích hình dạng quang phổ. 1. Introduction Landsat 7 was launched from Vandenberg airbase in California on April 15, 1999 with a Delta II missile. The satellite carries a sensor called ETM + (sensor for specialized map quality improvement). Like the OLI adjacent sensor, ETM + (Table 1) has 3 spectral channels in the visible region, 2 spectral channels in the near infrared spectrum, while the TIRS collects signals in two thermal infrared bands. Recent studies have demonstrated the potential of Landsat 5 / TM, ETM +, OLI for ocean color- related applications where its average spatial resolution (30m) provides detailed information (Pahlevan and Schott, 2013; Vanhellemont and Ruddick, 2014). Atmospheric correction algorithms have been developed to retrieve the remote sensing reflectance (Rrs( λ), where λ is the wavelength from the measured atmospheric peak radiation (TOA) signal by TM, ETM + or OLI (Franz et al, 2015; Pahlevan et al, 2017). Although there are differences in the sensitivity of the radiation measurements other sensors to the Landsat satellite stream, the evaluation of water parameters, such as suspended particulate matter concentration (SPM), disc depth Secchi, and colored dissolved organic matter (CDOM), from (Rrs(λ) are possible, mainly from empirical and regional approaches, with relatively good accuracy for with coastal waters (Vanhellemont and Ruddick, 2014; Concha and Schott, 2016; Kutser et al., 2016; Lymburner et al., 2016; Lee et al., 2016; Urbanski et al., 2016; Slonecker et al. , 2016; Li et al., 2017). Table 1. List of spectrum channels and corresponding resolutions of satellite Landsat-7/ ETM+ Many algorithms have been developed to detect and extract surface water from remote ISSN 2354-0575 Khoa học & Công nghệ - Số 27/Tháng 9 - 2020 Journal of Science and Technology 75 sensing technology in about two decades. For example, the Normalized Differential Vegetation Index (NDVI), developed in 1973 by Rouss et al., Is based on the red and infrared bands, which are used to distinguish the soil humus layer from clear water to chisel (Vermote and Saleous, 2007; Jawak et al., 2015). To determine the water pixels, the normalized difference water index (NDWI), which is based on the difference between reflectance in the green band, is heavily influenced by the water surface and the Near-Infrared reflectance Spectrometer (NIR) , was introduced by McFeeters (1996). However, this commonly used metric was unable to effectively suppress the signals from soil and water appearing in the image. Therefore, combine this with GIS (McFeeters, 2013) or the Modified Normalized Difference Water Index (MNDWI) by replacing the average infrared range with the near-infrared band (Xue). , 2006) to improve water surface detection in environments with complex information (images with different coatings layer). Due to the relatively good performance of the algorithm, as well as the full automation, the Fmask algorithm (The Function of the Mask) is often used to mask clouds, shadows and snow in Landsat images (Zhu and Woodcock, 2012). This algorithm is based on predefined thresholds of TOA reflectance values at specific ranges, calculates cloud probabilities, analyzes temperature, luminance, normalized Difference Vegetation Index (NDVI), and normalized differential snow index (NDSI), for a variety of objects (earth, water, shadow clouds, snow and clouds). Despite recent improvements by Fmask (Zhu et al., 2015), based on improvements on the Landsat 8 satellite and using a dynamic threshold instead of a fixed threshold for detecting clouds on water, some identifies Water pixel false positives are still limited especially in turbid waters (confused/ not recognized). Due to the high contribution of the Rayleigh component to the TOA signal, its contribution should therefore be eliminated in order to better characterize the spectral shape of the reflectance per object (Nordkvist et al., 2009). For these reasons, the current algorithm, called WiPE using atmospheric reflectance spectra peak have been corrected Rayleigh (ρrc(λ)) as input parameters. WIPE based on the combination of the common criteria are applied to (ρrc(λ)) to differentiate primarily pixels with pixel ground water, clouds and vegetation, and based on the application of analytical Hue -Saturation-Value or Brightness, HSV to optimize the detection of thin cloud and shadow pixels in the waters. 2. Research data Figure 1. Algorithm development collection image (blue) and algorithm test image (red) The Landsat-7 downloaded from USGS (https://landsat.usgs.gov/) based the different geographical locations including atmospheric conditions and environmental contrast (Figure 1). The entire data set has been divided into two different subgroups. The first group, made up of 196,738 different pixels (with 164,681 water pixels) extracted from 32 different images, was used to develop the algorithm. The second group consists of 12 different images for independent testing and algorithm validation. The wide spectral resolution range of water is reflected in the large variation of the Rayleigh corrected reflection spectrum, ρrc(λ), extracted on water pixels (Figure 2). Figure 2. Reflect spectrum of water pixels in algorithm development database ISSN 2354-0575 Journal of Science and Technology76 Khoa học & Công nghệ - Số 27/Tháng 9 - 2020 2.1. Eliminating the Influence of Rayleigh The TOA reflectance value, ρ TOA (λ), is developed as follows (Gordon, 1997): (1) which ρR(λ) is the reflectance due to multiple scattering of a pure atmospheric air molecules, ρa(λ) is the reflectivity of the many scattering by aerosols in a pure atmosphere, and ρRa(λ) stands for the term interacting in real atmospheres that contain both molecules and aerosol. T(λ) and t(λ) are the direct transmission and diffusion of the atmosphere, respectively. ρ g (λ), ρ wc (λ) and ρ w (λ) are the reflectivity coefficients due to sunlight, white light, and optically important components in water, respectively. The reflectivity of sunglint and whitecaps are related to wind speed and geometry (Wang and Bailey, 2001). The pixel value recorded by ETM + at the top of the atmosphere is a digital number, from which the TOA radiation, L TOA (λ), and reflectivity, ρ TOA (λ), are calculated using equations (2) and equation (3), respectively (Vanhellemont and Ruddick, 2014). (2) (3) where L TOA (λ) is the spectral radiation (in Wm-2sr- 1μm-1), ML(λ) the radiation multiplier ratio (i.e. gain) for each range, Qcal(λ) is the Level-1 pixel value in numerical form and AL(λ) is the rate of the radiation addition rate factor (ie deviation) for each band. d, F0 is the Earth-Sun distance in astronomical units and θ 0 is the average solar radiation of the band and the angle of the sun’s peak. The reflection coefficient due to Rayleigh scattering in atmosphere, ρR (λ), can be defined using the following equation (Gordon et al., 1988; Wang and King, 1997): (4) where τR(λ) is the thickness of Rayleigh scattering, pR(Θ) is a function of Rayleigh scattering phase at a scattering angle, Θ, θv is the angle of the celestial viewing angle and θ 0 the view of the sun. The optical thickness at atmospheric pressure P is calculated as follows: (5) where τ R0 (λ) is the optical thickness at standard atmospheric pressure P 0 is 1013.25 mbar. The Rayleigh pR(Θ) scattering phase function is defined as: (6) Terminology related to Θ- provides a contribution to backscatter of photons from atmosphere without interacting with the sea. The terminology related to Θ+ stands for the photon dispersed in the atmosphere towards the sea surface. These terminologies are calculated from solar sensor variable geometry as follows: (7) where Φ 0 and Φv are the sun and azimuth of the sensor; r(θ) is the Fresnel reflection coefficient for incident rays in the air at incident angle θ. A phase function Rayleigh, PR(Θ), of the scattering angle, Θ, and the Fresnel reflectance, r(θ), of the incident ray in the air at the incident angle is ditermied by: (8) (9) (10) The Rayleigh corrected reflectance is defined by Equation (11). (11) The Rayleigh scattering nomarlly contributes about 80% to 90% of the TOA reflectivity in the blue portion of spectrum for clear to turbid water (Gordon et al., 1988; IOCCG, 2010). 2.2. The Reflection Spectrum Database of Classified Objects All TOA reflectance coefficients and Rayleigh- corrected spectra are stored for each object including clouds on land and water (2430 pixels), shadows on land and water (2435 pixels), water (755344 pixels), vegetation (2838 pixels), bare soil (333 pixels) and buildings (340 pixels). All spectrum data of rrc(λ) for selected objects which are non-water is given in Figure 3. ISSN 2354-0575 Khoa học & Công nghệ - Số 27/Tháng 9 - 2020 Journal of Science and Technology 77 Figure 3. Rayleigh calibrated spectral data; (a) bare land, (b) buildings, (c) cloud, (d) vegetation Benefit of using ρrc(λ) instead of ρTOA(λ) to detect water pixels, which has typical reflection spectra, is illustrated in Figure 4. From clear water (Figure 4a) to very turbid water (Figure 4d), the reflection spectral shape (as well as amplitude) of ρrc(λ) has significantly different for various water types. Figure 4. Difference of reflectance spectrum ρrc(λ) and ρTOA(λ) for different water types 3. Developing WiPE Algorithm for Landsat-7/ ETM+ Images Typical spectrum examples of the ρrc(λ), which allows understanding different spectral criteria adopted in WiPE, are shown on five non- water objects in Figure 3. With a relatively large increasing ρrc(λ) from band 1 (blue) to band 5 (near infrared), vacant land and buildings have significant difference compared to other objects. Peaks in spectral channels 5 (near infrared) and 7 (shortwave infrared) are incorporated to a spectral shape which allows distincting vacant land and buildings from other objects. Similarlly, intensive increase of ρrc(λ) is observed between the red and near infrared channels for vegetation pixels; this is a remarkable feature for their identification. Based on relatively flat of spectral shape ρrc(λ) and a high signal level in the blue-red part of spectrum, it allows relatively well-defined clouds from other objects. Figure 5. Diagram of the WiPE algorithm for Landsat-7/ETM + sensor Figure 6. Relationship between the ratio ρrc(4)/ρrc(2) and the shortwave near infrared channel ρrc(7) For a Landsat-7/ETM+ image, the Rayleigh corrected reflection spectrum at shortwave infrared, ρrc(7), and the ratio between the near infrared channel and the channel green, ρrc(4)/ρrc(2), will help to classify all non-water objects excepted shadow objects. As above mentioned factors are manually identified through graphical analysis (Figure 6). The removed shadow objects are carried over to the HSV transformation steps (Figure. 5). The S and V values of the Rayleigh corrected spectrum are calculated as follows: ISSN 2354-0575 Journal of Science and Technology78 Khoa học & Công nghệ - Số 27/Tháng 9 - 2020 (12) (13) (14) 4. Applying WiPE Algorithm for Satellite Imagery Landsat-7/ETM+ Results in Table 2 are acquired based on four scenes (Figure 7) which are captured in downstream of Mekong River at different and random times. a) b) c) d) e) f) g) h) Figure 7. Results of LS7WiPE algorithm for downstream Mekong River Table 2. Average percentage difference between automatic (LS7WiPE) and manual (QGIS) water separation methods Figures Image name LS7WiPE (Pixel) QGIS (Pixel) MAPD (%) 7a LE07_L1TP_1240 53_20021207_201 70128_01_T1 28666001 28607521 0.2044 7c LE07_L1TP_1240 53_20030108_201 70126_01_T1 30359420 30351035 0.0276 7e LE07_L1TP_1240 53_20030313_201 70126_01_T1 21772494 21009847 3.6299 7g LE07_L1TP_1240 53_20030329_201 70125_01_T1 20948910 20945733 0.0151 5. Conclusions A new algorithm (WiPE and LS7WiPE) has been developed to evaluate water pixels for present optical sensors. Unlike the existing algorithm, WiPE based on on Rayleigh corrected reflectance, (𝜆); therefore, it is more sensitive to the spectral characteristics of various considered objects such as clouds, shadow clouds, vegetation, bare land, buildings, and water. WiPE generally shows very good performance for detecting water pixels in complex water environments. Acknowledgments This study is financially supported by the Code Senior Researcher program under contract number NVCC34.01/20-20. ISSN 2354-0575 Khoa học & Công nghệ - Số 27/Tháng 9 - 2020 Journal of Science and Technology 79 References [1]. Pahlevan N., Schott J. R.. Leveraging EO-1 to evaluate capability of new generation of Landsat sensors for coastal/inland water studies. IEEE Journal of Selected Topics in Applied Earth Observa- tions and Remote Sensing, 2013, Vol. 6, pp. 360-374. [2]. Vanhellemont Q., Ruddick K.. Turbid wakes associated with offshore wind turbines observed with Landsat 8. Remote Sensing of Environment, 2014, Vol. 145, pp. 105-115. [3]. Concha J. A. and J. R. Schott. Retrieval of color producing agents in Case 2 waters using Landsat 8. Remote sensing of Environment, 2016, 185, pp. 95-107. [4]. Kutser Tiit, Pascual Gema Casal, Barbosa Claudio, Paavel Birgot, Ferreira Renato, Carvalho Lino, Toming Kaire. Mapping inland water carbon content with Landsat 8 data. International Journal of Remote Sensing, 2016, 37, pp.13-2950. [5]. Lymburner, L., Elizabeth Botha, Erin Hestir, Janet Anstee, Stephen Sagar, Arnold Dekker, Tim Malthus. Landsat 8: Providing continuity and increased precision for measuring multi-decadal time series of total suspended matter. Remote Sensing of Environment, 2016, Vol. 185, pp. 108-118. [6]. Lee Z-P, Shaoling Shang, Lin Qi, Jing Yan, Gong Lin. A semi-analytical scheme to estimate Secchi-disk depth from Landsat-8 measurements. Remote Sensing of Environment, 2016, Vol. 177, pp. 101-106. [7]. Urbanski J-A, Agnieszka Wochna, Iwona Bubak, Waldemar Grzybowski, Katarzyna Lukawska- Matuszewska, Magda Łącka, Sylwia Śliwińska, Bożena Wojtasiewicz, Marek Zajączkowski. Application of Landsat 8 imagery to regional-scale assessment of lake water quality. International Journal of Applied Earth Observation and Geoinformation, 2016, Vol. 51, pp. 28-36. [8]. Slonecker E. T., Daniel K. J., Brian A. P. The new Landsat 8 potential for remote sensing of colored dissolved organic matter (CDOM). Marine Pollution Bulletin, 2016, Vol. 107, pp. 518-527. [9]. Li Y., Yunlin Zhang, Kun Shi, Yongqiang Zhou, Yibo Zhang, Xiaohan Liu, Yulong Guo. Spatiotemporal dynamics of chlorophyll-a in a large reservoir as derived from Landsat 8 OLI data: understanding its driving and restrictive factors. Environmental Science and Pollution Research, 2017, ISSN 0944-1344. [10]. S.K. McFeeters. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. International Journal of Remote Sensing, 1996, Vol. 17, pp. 1425-1432. [11]. McFeeters, S.K., Using the normalized difference water index (NDWI) within a geographic information system to detect swimming pools for mosquito abatement: a practical approach. Remote Sens., 2013, 5, pp. 3544–3561. [12]. Vermote, E.F., Saleous, N.Z., Operational atmospheric correction of MODIS visible to middle infrared land surface data in the case of an infinite Lambertian target. In: Qu, J.J. (Ed.), Earth Science Satellite Remote Sensing, Science and Instruments, 2006, vol. 1, Tsinghua Univ. Press, Beijing, pp. 123–153. [13]. Jawak, S.D., Kulkarni, K., Luis, A.J., A review on extraction of lakes from remotely sensed optical satellite data with a special focus on cryospheric lakes. Adv. Remote Sens, 2015, 4, pp. 196– 213. [14]. Xue, H., Modification of normalized difference water index (NDWI) to enhance open water features in remotely sensed imagery. Int. J. Remote Sens, 2006, 27 (14), pp. 3025–3033. [15]. Zhu Z., Woodcock C. E. Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sensing of Environment. 2012, Vol. 118, pp. 83-94. [16]. Zhu Z., Wang S., Woodcock C. E., Improvement and expansion of the Fmask algorithm: cloud, cloud shadow, and snow detection for Landsats 4-7, 8, and Sentinel-2 images. Remote Sensing of Environment, 2015, Vol. 159, pp. 269-277. ISSN 2354-0575 Journal of Science and Technology80 Khoa học & Công nghệ - Số 27/Tháng 9 - 2020 [17]. Nordkvist K., Loisel H., Gaurier L. D. Cloud masking of SeaWiFS images over coastal waters using spectral variability. Optics Express, 2009, Vol. 17, pp. 12246-12258. [18]. Gordon, H. R. Atmospheric correction of ocean color imagery in the Earth Observing System era. Journal of Geophysical Research, 1997, 102, pp. 17081-17106. [19]. Wang M., and S. W. Bailey. Correction of sun glint contamination on the SeaWiFS ocean and atmosphere-products. Applied Optics, 2001, 40, pp. 4790-4798. [20]. Gordon H. R., Brown J. W., Evans R. H. Exact Rayleigh scattering calculations for use with the Nimbus-7 coastal zone color scanner. Applied Optics, 1988, Vol. 27, pp. 862-871. [21]. Wang M., Michael D. King. Correction of Rayleigh scattering effects in cloud optical thickness retrievals. Journal of Geophysical Research, 1997, Vol. 102, pp. 25915-25926. THUẬT TOÁN TỰ ĐỘNG TÁCH MẶT NƯỚC CHO ẢNH VỆ TINH LANDSAT-7/ETM+ CHO VÙNG NƯỚC VEN BỜ VÀ TRONG ĐẤT LIỀN Tóm tắt: Việc xác định các pixel mặt nước trên các vùng nước tự nhiên là giai đoạn quan trọng trước khi áp dụng các thuật toán dành riêng cho việc tính toán các thông số môi trường nước từ ảnh vệ tinh viễn thám. Một số thuật toán tồn tại nhưng hiệu suất của chúng không đạt yêu cầu, đặc biệt là trên vùng nước đục, nơi các pixel không có mây đôi khi được phân loại là mây hoặc đất, dẫn đến mất dữ liệu. Điều này đặc biệt quan trọng đối với các ảnh vệ tinh viễn thám có độ phân giải không gian cao, chẳng hạn như các quan