MODIFIED OPTIMIZATION WATER INDEX ( MOWI ) FOR LANDSAT-8 OLI / TIRS

Water is one of the most important resources that essential need for human life. Due to population growth and increasing need of human to water, proper management of water resources will be one of the serious challenges of next decades. Remote sensing data is the best way to the management of water resources due time and cost effectiveness over a greater range of temporal and spatial scales. Between many kinds of satellite data, from SAR to optic or from high resolution to low resolution, Landsat imagery is more interesting data for water detection and management of earth surface water. Landsat8 OLI/TIRS is the newest version of Landsat satellite series. In this paper, we investigated the full spectral potential of Landsat8 for water detection. It is developed many kinds of methods for this purpose that index based methods have some advantages than other methods. Pervious indices just use a limited number of spectral band. In this paper, Modified Optimization Water Index (MOWI) defined by consideration of a linear combination of bands that each coefficient of bands calculated by particle swarm algorithm. The result shows that modified optimization water index (MOWI) has a proper performance on different condition like cloud, cloud shadow and mountain shadow. * Corresponding author


INTRODUCTION
Water is one of most important earth resources that about 71 percent of the Earth's surface is covered by water (Williams 2014).Rapid population increase and unplanned urbanization leading to decrease of water bodies.Monitoring and proper management of water resources such rivers, lakes as natural reservoirs and dams as a man-made reservoir are essential to human health, society, agriculture, global carbon cycle and climate variations (Brezonik, Menken et al. 2005, Prasad, Rajan et al. 2009).Remote sensing data has been widely used for water detection in last 3 decades rather than ground-based methods with the advent and development of remote sensing sensors and methods due to a time and cost effectiveness over a greater range of temporal and spatial scales (Wang, Ruan et al. 2011).Therefore, water detection is an interesting field of experts and researchers of remote sensing and photogrammetry.
Landsat imagery is one of the most wildly used remote sensing data that applied in most previous studies for water body detection or water change detection.The Landsat series of satellites have many sensor and generation.Multispectral Scanner System (MSS) for landsat-1 to landsat-3 (1972-1983), Thematic Mapper (TM) for landsat-4 to landsat-5 (1982-2012) and Enhanced Thematic Mapper Plus (ETM+) for landsat-7 (1999-now) have been used for many application especially water index and water body detection (McFeeters 1996, Xu 2006).The newest generation of Landsat series of satellites is landsat-8 that lunched on 2013.Operational Land Imager (OLI) sensor of landsat-8 has 12-bit pixel value rather than the 8-bit pixel value of ETM+ images that lead to higher quality and improve three times better signal to noise ratio (SNR) than ETM+ (Irons, Dwyer et al. 2012).
There are many kinds of algorithm have been adopted for water detection especially on Landsat data that categorized into four main groups: 1-classification and pattern recognition methods that include supervised (Tulbure and Broich 2013) and unsupervised methods (Ko, Kim et al. 2015), 2-spectral unmixing (Sethre, Rundquist et al. 2005), 3-single-band thresholding (Klein, Dietz et al. 2014), 4-the spectral water index (Ji, Zhang et al. 2009).Classification methods have a better performance than thresholding methods.If images consist of complex topologies such as mountain shadows, roads, and urban areas, high false classification rate may be achieved in water body detection process (Ko, Kim et al. 2015).Also, classification methods are highly dependent on human and have some difficulty (Ouma and Tateishi. 2006).Index-based methods can detect water body more accurately, quickly and easily than classification methods and does not need any prior knowledge (Li, Du et al. 2013), especially on low-resolution images and single-class (water) study.
There are many kinds of water indices in past studies.One of popular water index is the normalized difference water index (NDWI) (McFeeters 1996).The other one is the modified NDWI (Xu 2006) that introduced to overcome the inseparability of built up areas in NDWI.Automated water extraction index (AWEI) introduced for better result achieving in an area by shadow and dark surface on Landsat TM.AWEInsh and AWEIsh are used for an area by urban background and area by shadow respectively (Feyisa, Meilby et al. 2014).Water Ratio Index (WRI) is another widely used water index (Shen and Li 2010).Amare Sisay (2016), used NDWI for MSS image data due lacks mid-infrared band (MIR) and AWEI for TM, ETM+, and OLI_TIRS for change detection of central rift valley region of Ethiopia (Sisay 2016).DU et al. (2014), tested three NDWI models include NDWIO5,3, NDWIO6,3, NDWIO7,3 based on reflectance value that result show the better accuracy of NDWIO6,3 than two other models (Du, Li et al. 2014).Liu et al. (2016), indicated that NDWI3,5 and NDWI5,6 have better performance by reflectance images whereas NDWI3,6 has better performance by DN value images (Liu, Yao et al. 2016).Blackmore (2016), showed the effectiveness of NDWI and MNDWI for open water and surface moisture detection.Also, results of this study indicated that NDWI is more sensitive to an area with vegetation (Blackmore 2016).Xie et al. (2016), studied about clear water, turbid water, and eutrophic water.The results indicated that AWEIsh, NDWI4,7 and NDWI3,7 with accuracies of 98.55%, 95.50%, and 96.61% have the highest accuracy to clear water, turbid water, and eutrophic water, respectively (Xie, Luo et al. 2016).
It should be noted that all of the above mentioned spectral indices only use a limited number of bands that may lead to the poor result for pixels that contaminated by ice, snow or cloud.In this paper, a modified optimization water index (MOWI) is proposed to use the full spectral potential of landsat-8 OLI/TIRS as the newest generation of Landsat series of satellites in water detection and reduce shadow effects, cloud effects and other disturbing factors.The proposed method can be considered as classification method or index-base method that we have an index-based view in this paper.

WATER INDICES
In this section, an overview of the water indices is presented.By considering high reflectance of green band and low reflectance of near infrared band for water, normalized difference water index (NDWI) defined as bellow (McFeeters 1996): Modified normalized difference water index (MNDWI) is achieved with replacing infrared by shortwave infrared that its wavelength is [1.57-1.65]micrometers (SWIR1).MNDWI defined as bellow (Ji, Geng et al. 2015): Water ratio index is another water index that defined as bellow (Rokni, Ahmad et al. 2014, Sisay 2016):

THE PROPOSED METHOD
All of the previous indices calculated by the commentary of spectral curvature but in this paper, another approach has been adopted that use combination of bands for water detection under a linear combination of bands.The coefficient of each band is determined by the optimization algorithm.This linear combination defined as bellow: 1 (5) Proposed method can be adopted at two strategies: 1-Classification-based: in this strategy, coefficients calculated by some test and some terrain pixel.Then water body or surface can be detected by applying these coefficients.These coefficients are proper just for this image.2-Index-based: in this strategy, coefficients calculated on one image by test and obtained water index.Then we can apply this index to other images.In this study, the index-based strategy has been used for water detection that used all bands instead of using a limited number of bands.The flowchart of the proposed method is presented in Figure 1.

Study area and Data source
Landsat-8 is the newest generation of Landsat series of satellites that is the most interesting source for water detection studies.In order to reach the final aim of this study, four Landsat 8 OLI/TIRS images were collected from the US Geological Survey (USGS) Global Visualization Viewer (collection 1 level-1).All of these four case studies are in Iran and includes Zarivar Lake, Kazemi Dam, Gotvand Dam and Karun-1 Dam.These images have a different condition like cloud and shadow that are disturbing factor for water detection.In this paper, we wanted to use all potential of spectral space for water index definition.So, 10 bands of landsat-8 had been used to determined water index.Table 1 presents the specifications of Landsat 8 OLI/TIRS images (Rokni, Ahmad et al. 2014, Sisay 2016).

Preprocessing
Preprocessing is one of the most important steps for each photogrammetry and remote sensing analysis.In this paper, the original DN values had to be converted into radiance and then into reflectance.

Optimization
Particle swarm optimization (PSO) is a population based metaheuristic algorithm that has a simple programming, high run speed and high convergence rate that proposed in 1995 by Kennedy and Eberhart (Eberhart and Kennedy 1995).Here, PSO is used to find optimum coefficients for each band for have a proper water index.The overall accuracy is considered as objective function and variables (coefficients) domain is considered in [-10, 10].The coefficients of the index (modified optimization water index (MOWI)) are calculated on Karun-1 Dam due to having water, shadow, and cloud.

Evaluation
After calculated coefficients and modified optimization water index on Karun-1 Dam, MOWI evaluated on three other case studies by the recall, Precision (P), f-score, overall accuracy (OA) and kappa coefficient (K).

EXPERIMENTAL RESULTS AND DISCUSSIONS
Coefficients of each band that calculated by particle swarm optimization are as bellow.In the case study of Karun-1 Dam, MOWI has the best performance by 99.7085, 99.9301 and 0.9967 for f-score, overall accuracy, and kappa coefficient respectively.The result of this case study shows the problem of MNDWI and WRI by cloud and its shadow.Also, AWEIsh has a little problem for narrow water detection especially at down-right and NDWI has a little weak result in water detection at down-left.
In the case study of Gotvand Dam, AWEIsh has the best performance by 99.9970, 99.9978 and 0.9999 for f-score, overall accuracy, and kappa coefficient respectively.However, this index has less noise than other indices but has a little problem for narrow water detection at middle-top and middleleft.After AWEIsh, MOWI has the best performance in this case study by 99.9881, 99.9914 and 0.9998 for f-score, overall accuracy, and kappa coefficient respectively that has a little difference by AWEIsh.
In the case study of Kazemi Dam, MNDWI and WRI have the best result but MOWI and NDWI performance have the little

CONCLUSION
Water is one of the important resources that should be managed properly for needs of human life in future.One of best tools for water resource management is remote sensing data and techniques.Also, water detection is one of interest subjects of photogrammetry and remote sensing researcher.Between many kinds of satellite data, Landsat imagery is the more interesting data for water detection especially Landsat8 OLI/TIRS that is the newest version of Landsat satellite series.In this paper, we investigated the full spectral potential of Landsat8 to calculate the water index by consideration of the linear combination of bands.Particle swarm optimization had been sued for calculation of each band coefficient.The result showed that modified optimization water index (MOWI) has a proper performance on different condition like cloud, cloud shadow and mountain shadow.
index (AWEI) use four bands unlike two bands of NDWI and MNDWI.This index defined as bellow(Ji, Geng et al. 2015):

Figure 1 .
Figure 1.Schema of the proposed method

Table 4 .
Result of each index on each case studyThe International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W4, 2017 Tehran's Joint ISPRS Conferences of GI Research, SMPR and EOEC 2017, 7-10 October 2017, Tehran, Iran difference by MNDWI and WRI.Also, AWEIsh the lowest accuracy.Results of Zarivar Lake show the better performance of MOWI by 96.4250, 98.1862 and 0.9521 for f-score, overall accuracy, and kappa coefficient respectively.Also, WRI has the poor result on this case study.Results of four subsets show that MNDWI and WRI have the problem by cloud and its shadow, and AEWI has a little problem for narrow water detection.MOWI has the best result in two case study and has the little difference by best results of two other case study result.Totally, MOWI uses all potential of landsat-8 images for water detection that has the desirable results on the different case study by different condition such as cloud, its shadow and mountain shadow.