144.2017

J.Natn.Sci.Foundation Sri Lanka 2018 46 (3): 329 - 339 DOI: http://dx.doi.org/10.4038/jnsfsr.v46i3.8485

RESEARCH ARTICLE

An automated image dehazing method for flood detection to improve flood alert monitoring system

Chien-Hao Tseng1, Lun-Chi Chen1, Jyh-Horng Wu1, Fang-Pang Lin1 and Ruey-Kai Sheu2* 1 National Center for High-Performance Computing, National Applied Research Laboratories, Taiwan (ROC). 2 Department of Computer Science, Tung Hai University, Taichung, Taiwan (ROC).

Revised: 18 January 2018; Accepted: 23 March 2018

Abstract: Flood hazard prevention and mitigation is an the main causes of flooding, and heavy rain is the most emergent environmental problem. Traditional flood monitoring dominant among these factors. Prolonged rainfall causes devices operated under adverse environmental conditions are rivers to overflow, consequently flooding the surrounding typically influenced by changes in weather conditions such as areas. Although video surveillance systems can be used haze, fog and rain. Consequently, the video images obtained to forecast rainfall or track storm paths precisely, the from such devices are often blurred or damaged, increasing availability of real-time monitoring data such as water the possibility of erroneous assessments in hazard mitigation flow, flood level, and water level is essential in making processes. To ensure the efficient use of image analysis reasonable decisions on necessary flood prevention technology to improve degraded images captured under hazy measures. weather conditions, this study proposes an automated single image dehazing method for flood monitoring. This method is based on the dark channel prior for the removal of haze from Recently, video surveillance systems have been a single image. The concept of the dark channel prior is that effectively used for monitoring flood events. Flood most local patches in haze-free outdoor images contain some monitoring at embankments, bridges, levees and dams pixels having extremely low intensities in at least one colour is essential for flood prevention. Various methods can channel. When this dark channel prior is used, the thickness of be used to monitor and prevent flood events. One of haze in the image can be directly estimated and a high-quality the most challenging problems in flood response is the haze-free image can be obtained. The proposed method can precise localisation of flood risk areas, which is typically be used to accurately improve flood detection and monitoring conducted using early warning systems (EWSs) for results. The ability to detect and remove haze from a single flood prevention and disaster management. EWSs image is a crucial function when applying automated computer have been used for monitoring flooding by employing vision to disaster-monitoring applications. The experimental remote sensing technologies such as satellite imaging results show that the proposed method can efficiently alleviate and electronic sensors installed near rivers and seaports the degradation of surveillance images and effectively identify (Basha et al., 2008; Krzhizhanovskaya et al., 2011). Flood flooded regions in particular areas. monitoring using real-time sensors is a non-structural flood control measure. Losses caused by flooding can Keywords: Flood monitoring, hazard mitigation, haze imaging be mitigated or prevented by using measures such as model, image dehazing. flood monitoring, forecasting, simulation, evaluation and analysis (Chen, 1990). One non-structural measure, the INTRODUCTION integration of wireless sensors with web-based decision support systems, has been influential in monitoring, Flooding is a major disaster occurring in various parts controlling, and assessing natural disasters, particularly of the world. Heavy rain, hurricanes, and tsunamis are flood disasters (Zhang et al., 2002; Sapphaisal, 2007).

* Corresponding author ([email protected]; https://orcid.org/0000-0002-3014-8095)

This article is published under the Creative Commons CC-BY-ND License (http://creativecommons.org/licenses/by-nd/4.0/). This license permits use, distribution and reproduction, commercial and non-commercial, provided that the original work is properly cited and is not changed in anyway. 330 Chien-Hao Tseng et al.

Effectively implementing a flood monitoring and warning Wijeyaratne, 2013) from videos or sensors has recently system is imperative because such a system requires the received considerable attention. In the current study, we availability of reliable and related information. Flooding focused on haze detection and removal from surveillance has recently been studied using various considerations images. Surveillance images captured from optical and methodologies such as image processing technology devices are usually degraded by turbid media such as (Rani et al., 2011; Selvi & Sathya, 2014; Wu et al., 2015), haze, smoke and fog. Haze is the most common problem wireless sensor networks (Chang & Guo, 2006; Hughes in outdoor scenes because of atmospheric conditions. et al., 2006), and middleware embedded systems (Shukla When the depth of haze in a hazy image is unknown, & Pandey, 2014). A decision-theoretic methodology for dehazing the image becomes a challenge. Furthermore, modelling and evaluating forecast response systems can if the available input is only a single hazy image, the be combined with loss functions and prior knowledge problem is under-constrained. Hence, most traditional about actual river water levels (Krzysztofowicz & Davis, dehazing approaches (Narasimhan & Nayar, 2003; Namer 1983). Albanese et al. (2008) have reported that theories et al., 2009) entail using multiple hazy images as inputs derived from mathematical-morphology-based image or additional prior knowledge. Moreover, polarisation- processing methods can be used to develop automated based methods have been proposed for removing haze detection systems that can provide flood warnings. These effects from two or more images taken under different studies provide great insight into the development of polarisation conditions (Schechner et al., 2003; Shwartz flood forecasting and modelling systems by using remote et al., 2006). Nevertheless, taking multiple images of the sensing data and digital image processing techniques. same scene is usually impractical in real-time monitoring Flood alarm systems are one of the most widely used system applications. He et al. (2010) proposed an infrastructures for creating flood monitoring and effective image prior, called the dark channel prior, for warning systems. Closed-circuit television (also called removing haze from a single image. The key concept of video surveillance) systems are another example of such the proposed prior is that most local patches in outdoor infrastructures. haze-free images contain certain pixels having extremely low intensity values in at least one colour channel. Flood monitoring generally involves using on-site water-level measurement facilities, such as rainfall In this paper, inspired by a single-image-based observation stations, water level observation stations, dehazing framework based on the dark channel prior and meteorological stations. On-site stations can methodology, an image -based flood alarm module directly measure the water or rainfall level and provide (IFAM) that can be integrated into an automated instant notifications. However, using sensors to directly single image dehazing algorithm for flood monitoring measure the water level is particularly restricted by the is proposed. The IFAM is the foundation of the image sensor installation location and necessity of frequent analysis technique employed in this study to detect and maintenance. Furthermore, this direct measurement identify flood conditions in real-time video images. approach has the disadvantage of returning only In foggy weather during typhoon, the fuzzy images water-level information and not visual evidence for caused by fog or haze have a negative influence on the assessment. Therefore, the integration of image analysis recognition and detection of water stage at the backend, techniques with flood alarm systems is vital for flood which render the warning process less effective. Thus, this disaster prevention. Although real-time videos of river paper proposes an image dehazing technique to improve surroundings or urban areas can be easily obtained from the image quality, in order to elevate the accuracy and existing video surveillance systems, various uncertainties liability of image analysis on water stage detection. The in the outdoor environment may affect the video quality. module is connected to outdoor surveillance cameras Therefore, video surveillance systems are influenced by installed around high-risk rivers and other places. haze, fog, rain and other atmospheric conditions, which Moreover, an automatic single image dehazing method increase the possibility of erroneous assessment. was employed based on the dark channel prior to improve the flood detection and monitoring results. The dark The removal of weather effects such as haze (Gibson channel prior is based on the statistics of outdoor haze- et al., 2012; Guo et al., 2014; Zhai & Ji, 2015; Patel & free images. The experimental results indicated that the Nakrani, 2016), fog (Chen et al., 2013; Tan et al., 2014), proposed method can be used to detect and remove haze rain (Barnum et al., 2010; Kang et al., 2012), and air from a single image and effectively identify flooding in pollution (Abeyratne & Ileperuma, 2006; Attanayaka & particular areas.

September 2018 Journal of the National Science Foundation of Sri Lanka 46(3) Flood monitoring system using dehazing method 331

Figure 1: Schematic of the haze imaging model applied to a natural scenario with haze Figure 1: Schematic of the haze imaging model applied to a naturalI x)( scenario= J with )( xtx haze)( + A 1( − xt ) )(

containx)( pixels )( havingxtx )( extremely1( xt ) low )( intensities in at least METHODOLOGY I = J + A − one colour channel. Consider an arbitrary image J; its dark channel dark can be expressed as follows:c  Image degradation model and haziness analysis J x)( = min min J y)( y Ω∈ x)(  ∈{} rc g ,, b 

In 1976, McCartney proposed a haze imaging model dark  c  J x)( = min min J y)( (also called image degradation model) comprising a y Ω∈ x)(  ∈{} rc g ,, b  direct attenuation model and an air light model. The ...(2) direct attenuation model describes the scene radiance  Ic y)(  where J c is a colourxt )( =1 channel− w min of J,min Ω(x) is a local patch and its decay in the medium, whereas the air light model y Ω∈ x)(  c Ac  is composed of previously scattered light and it describes centred at x, y is the index of a pixel in Ω(x), and c is one of the three colour channels in the RGB (red, green and the shift in the scene colours. As illustrated in Figure 1,  Ic y)(  haze is a turbid medium (e.g., dust particles, tiny water blue)xt )( =space.1− w min min  y Ω∈ x)( c c droplets, and molecules) in the atmosphere and degrades  A  outdoor images because of atmospheric absorption and When the dark channelxI )( prior− A is combined with the J x)( = + A air light scattering. The haze imaging model can be haze imaging model,max( the valuext ) ,( tof) air light A can be expressed as follows: directly estimated to recover a high-quality0 haze-free image. According to the concept of a dark channel, if J is xI )( − A Figure 2: Sketch of the proposed...(1) flood-alertanJ outdoorx)( = monitoring haze-free systimage,em+ A except for the sky region, the intensity max(of J’s darkxt ) ,( channelt0 ) is low and tends to approach dark whereI =J I is the observed + A1 intensity −  and the input hazy zero (i.e., J ˆ → 0 ),1 whichn is called the dark channel image, J is the scene radiance and the restored haze- prior. In this study,f x)( = the ∑hazeK H imaging(x − xi )model and dark n i=1 free image, A is the global atmospheric light that can channel prior were used to remove the effect of haze be obtained using the dark channel prior, and t is the from video images. The low intensity in the dark channel n medium transmission describing the portion of the isˆ mainly1 attributed to three properties: (1) shadows, f x)( = ∑ K H (x − xi ) light that reaches the camera. The major goal of single (2) colourfuln i= 1objects or surfaces and (3) dark objects or − 1 2/ image dehazing is to recover a haze-free image J, A surfaces. K x)( = H K (H − 2/1 x) H and t from the input image I. In this model, the term J(x)t(x) denotes direct attenuation, and the term A(1-t(x)) Single image dehazing based on dark channel prior n  2  is called air light. Direct attenuation denotes the scene x − xi − 1 2/ x exp  radiance and its decay in the medium, whereas the air HeK et al.x)( (2010)= H proposedK (H a∑− single2/1 xi ) image dehazingh  method H i=1   light results from previously scattered light and leads based on the darkm .Kh channelx)( = prior. This method is based on n  2  to a shift in the scene colours. This haze image model the assumption that the dark channel xof− ax hazyi image can n  exp2  has been employed in numerous studies involving single approximate the haze density;x∑− x hence,i  this study also used x exp i=1  h  image dehazing (Fattal, 2008; Zhang et al., 2010; Xiao the dark channel∑ toi estimate the regions with the highest i=1 h & Gan, 2012). hazem densitiesx)( in the surveillance images. The top 0.1 % .Kh = 2 brightest pixelsn in dark channelx − x locations of a hazy image exp i  The dark channel prior is a type of statistic used in were used, which∑ are usually most haze-opaque. Among i h processing haze-free outdoor images. It is based on the these pixels, pixels=1 with the highest intensity values in observation that most local patches in haze-free images the hazy image are considered as atmospheric light. m (x k ) x k+1 = x k + m (x k ) ∇fˆ x)( = 0 ,Kh i i i ,Kh i ,Kh Journal of the National Science Foundation of Sri Lanka 46(3) September 2018

m (x k ) x k+1 = x k + m (x k ) ∇fˆ x)( = 0 ,Kh i i i ,Kh i ,Kh I x)( = J )( xtx )( + A 1( − xt ) )(

dark  c  J x)( = min min J y)( 332 y Ω∈ x)(  ∈{} rc g ,, b  Chien-Hao Tseng et al.

Once the atmospheric light is known, in a noise-free The proposed image-based flood alarm framework image, the transmission map can be derived as follows:  I y)(  The IFAM is a flood alert system that facilitates proactive c monitoring, assessment, and early response to rising xt )( =1− w min min c  y Ω∈ x)(  c A  ...(3) water levels and associated inundations in near real-time. The IFAM receives video streams from digital camera where w (0 < w ≤ 1) is a scaling parameter used to sensors installed around rivers. The video streams are Imaintainx)( = J an )( xtx extremely)( + A 1( − smallxt ) )( amount of haze for a decomposed into JPEG images and filtered for image distant object,xI and)( − Ac represents a specific colour channel enhancement. The module then calculates water levels (i.e.,J x )( the= red, green or blue+ A channel of the corresponding and assesses the corresponding flood risks. Figure 2 parameter).max( Thus,xt )an ,( t0 estimate) for the transmission map illustrates the proposed flood alert monitoring system. candark be obtained by simply subtractingc  the dark channel The IFAM combines on-site real-time video images with Jof the normalisedx)( = min imagemin fromJ equationy)( (3). y Ω∈ x)(  ∈{} rc g ,, b  a backend image processing module to conduct near real-time river overflow and ground inundation analysis. After the ninitial haze estimate is obtained, a refinement After the water overflow range is calculated by using ˆ 1 processf x)( = is required∑ K H ( tox −suppressxi ) halo artifacts. Therefore, an image processing module, the system automatically t(x) is refinedn i=1 using the matting Laplacian matrix (Levin  I y)(  provides flood alarms if the overflow range exceeds the etxt al.)( =, 12008).− w min This minrefinementc  process provides visually preset warning ranges. satisfactory yresults,Ω∈ x)(  cand Ait cwas thus used in this study. Finally, the dehazed image is recovered by a simple The proposed IFAM-based flood alert monitoring inversion of equation− 1 2/ (1) −and2/1 solving for J. The haze-free system involves two major steps: image dehazing and K H x)( = H K (H x) image J can be recovered as follows: detection of flooded areas. The image dehazing process xI )( − A is based on the dark channel prior. This prior can be J x)( n A 2 J = + x − x  ...(4) used to directly estimate the thickness of the haze, and max( xt x) ,( texp0 )  i  ∑ i   a high-quality haze-free image can be obtained after i=1  h  wherem tx)( denotes a lower bound of t(x) used to preserve this estimation. In the detection of flooded areas, image .Kh 0 = 2 n   analysis technology is used to evaluate the water level a small amount of hazex − x iin extremely dense haze ∑exp in relation to the preset overflow range. The module regions. A ntypical value ofh t0 is 0.1. Because the scene ˆ 1 i=1   radiancef x)( = is∑ usuallyK H (x − notxi ) as bright as atmospheric light, then determines whether to send a warning message the surveillancen i=1 images appear dim after haze removal. depending on the calculated results. Details on the Therefore, to identifyFigure flooded 1: Schematic areas, the of exposurethe haze imaging of J(x) modeloperating applied to mechanisms a natural scenario of these with hazemodules are provided in was increased. the next section.

− 1 2/ − 2/1 K H x)( = H K (H x) k k +1 k k ˆ m ,Kh (xi ) xi = xi + m ,Kh (xi ) ∇f ,Kh x)( = 0 2 n  x − x  x exp i  ∑ i   i=1 h m x)(   .Kh = 2 n  x − x  exp i  ∑   i h =1  

m (x k ) x k+1 = x k + m (x k ) ∇fˆ x)( = 0 ,Kh i i i ,Kh i ,Kh

Figure 2: Sketch of the proposed flood-alert monitoring system Figure 2: Sketch of the proposed flood-alert monitoring system

September 2018 Journal of the National Science Foundation of Sri Lanka 46(3) 1. In foggy weather during typhoon, the fuzzy images caused by fog or haze have a negative influence on the recognition and detection of water stage at the Flood monitoring system using dehazing method 333 backend, which render the warning process less effective. Identifying flooded areas by using image processing ThereforeI x)( = J, in )( thisxtx )( study,+ A 1( region-based− xt ) )( image identification module methods were developed by combining the mean-shift (MS) and region growing (RegGro) algorithms. 2. The method is able to utilize image prior through a technique called dark In general, in a hazy image, the haze density is higher in deeper regions and lower in shallower regions. Tochannel The prior. MS algorithm is a powerful nonparametric clusteringdark algorithm involving nonparametricc  estimation estimate the haze density and remove the haze, the initial J x)( = min  min J y)(  haze transmission was first estimated by using the dark of probability ydensityΩ∈ x)(  ∈functions,{} rc g ,, b and it can be used to estimate the local density gradients of similar pixels. The channel prior. Furthermore, the atmospheric light values I x)( = J )( xtx )( + A 1( − xt ) )( are automatically estimated, and the sky regions3. areFurthermore, gradient estimates utilizing are thisused image-dehazingin an iterative procedure techniqu to e improves the quality of determine the peaks in the local density. In the iterative accurately segmented from the image according to theirimages, and can further assist the backend module to improve the effectiveness distinctive features. Therefore, the dark channel prior is stage of the MS procedure, each pixel is assigned an MS  Ic y)(  effective in recovering saturated colours and extracting pointxt )( =M1(x−i ),w whichmin is initialisedmin to coincide with the pixel. of recognition. y darkx)( c c  c  low-contrast objects. The resulting transmission maps A multivariate JkernelΩ∈  xdensity)( = minA estimator min withJ a bandwidthy)( y Ω∈ x)(  ∈{} rc g ,, b  are reasonable. Next, a bilateral filter was applied to matrix H computed at point x is derived as follows: obtain a refined haze transmission and to smoothen its n ˆ 1 small-scale textures. Finally, the haze-free image can be f (x) = ∑ K H (x − xi ) obtained by multiplying the albedo by the atmospheric n i=1 ...(5) xI )( − A  I y)(  light according to the haze imaging model. Fattal (2008) J x)( = xt )( 1 +wAmin min c  where themax( n = dataxt )= ,( pointst )− and x representc a sample with proposed an albedo estimation-based method for single 0 y Ω∈ x)( i c A  an unknown density f. image dehazing. The method is able to utilise image prior through a technique called dark channel prior. Figure 3 depicts a flowchart of the described process. ...(6) 1 n xI )( − A fˆ x)( = ∑ JKx)( ⁄( =x − x ) ⁄  + A where K(=x) is|H a| dH-dimensionalHmax(i xt )kernel ,( t ) function satisfying n i=1 0 Surveillance image grabbed the conditions for asymptotic unbiasedness, consistency, by digital camera and uniform consistency of the gradient of the density estimate. The term H is a symmetric positive definite

d × d bandwidth matrix. n Determine dark channel of image −ˆ 1 2/ 1 − 2/1 K H x)( = H f x)( K=(H∑ KxH)(x − xi ) If a Gaussian kerneln isi=1 used, the MS vector becomes 2 n  x − x  Estimating the global atmospheric light x exp i  ∑ i   i=1  h  m x)( − 1 2/ − 2/1 .Kh = K x)( = H 2K (H x) n H  x − x  Refine the transmission map ∑exp i  by using soft mapping  h  2 i=1  n   x − x  ...(7) x exp i  ∑ i   i=1 h m x)(   Recovering the scene radiation The term .Kh =is proportional to the2 normalised n  x − x  density gradient and is alwaysexp orientedi towards the ∑   maximal density increase. i h =1   Output haze-free image The MS algorithm involvs the following steps: m (x k ) x k+1 = x k + m (x k ) ∇fˆ x)( = 0 ,Kh i i i ,Kh i ,Kh Compute the MS vector for point m (x k ) Step 1: h,K i according to equation (7). Figure 3:Figure Flowchart 3: Flowchart of the of haze the hazeremoval removal method method forfor surveillancesurveillance images images Step 2: Updatem the(x currentk ) position: x k+1 = x k + m (x k ) ∇fˆ x)( = 0 ,Kh i i i ,Kh i ,Kh

k+1 k k xi = xi + mh,K (xi ) . According to the properties of the resulting haze-free Step 3: Repeat steps 1 and 2 until convergence; image, it was determined that region-based image specifically, repeat these steps until the classification methods are more suitable for flood region gradient of the density functions at the point is ˆ detection compared with pure pixel-based methods. zero (i.e., ∇f h,K (x) = 0 ).

Journal of the National Science Foundation of Sri Lanka 46(3) September 2018

Figure 4: Surveillance image sequence (from top to bottom, left to right) for the Nan-Hu No. 2 bridge dataset Surveillance image grabbed by digital camera

Determine dark channel of image

Estimating the global atmospheric light 334 Chien-Hao Tseng et al.

The RegGro algorithm is a simple region-based image Refine the transmissionRESULTS map AND DISCUSSION segmentation algorithm. This algorithm groupsby pixelsusing soft mapping into meaningful regions, starting from a specific seed In this study, the proposed IFAM-based flood alert pixel and spreading to neighbours that satisfy the growing monitoring system was employed to identify flooded rule. The fundamental disadvantage of intensity-basedRecovering the sceneareas. radiation All surveillance images of the flooded areas region segmentation is that intensity alone provides captured from streaming videos were evaluated. no spatial information and the threshold criterion is a Although the examples that this paper demonstrates can single value or a set of grey levels. Hence, in the RegGro be conducted by human eye, the large dataset will cause algorithm, a compact dynamic mean intensityOutput is used haze-free tiredness image to further influence the accuracy. Therefore, with a settled threshold (intensity distance) to implement this study can achieve the effectiveness of automated the growing rule. warning by automatically detecting the water stage. Figure 3: Flowchart of the haze removal method for surveillance images

Figure 4: Surveillance image sequence (from top to bottom, left to right) for the Nan-Hu No. 2 Figure 4: Surveillance image sequence (from top to bottom, left to right) for the Nan-Hu No. 2 bridge dataset

bridge dataset

Figure 5: FigureSurveillance 5: Surveillance image image sequences sequences (from (from top to top bottom, to leftbottom, to right) forleft the to Ma-Ling-Keng right) for Riverthe datasetMa-Ling-Keng

SeptemberRiver dataset2018 Journal of the National Science Foundation of Sri Lanka 46(3)

(a) (b) (c)

Figure 6: Image dehazing results: (a) original hazy images, (b) estimated transmission map, and (c)

restored haze-free images Flood monitoring system using dehazing method 335

Furthermore, utilising this image-dehasing technique The dehazed images were analysed using the mentioned improves the quality of images, and can further assist image analysis techniques to evaluate the water level of the backend module to improve the effectiveness the current overflow range. It was determined that the of recognition. All images were in JPEG format, region-based image identification methods are highly exhibiting a 24-bit colour depth and a spatial resolution suitable for flood region detection. The first step in image of 352 × 240. Field experiments were conducted using processing was to convert the colour space from the RGB data from the Nan-Hu No. 2 Bridge and Ma-Ling-Keng domain into a hue-saturation-value (HSV) domain. The River in Taiwan. The Nan-Hu No. 2 Bridge data were HSV domain is preferred because it is more perceptually recorded during the Typhoon Sinlaku, which induced relevant than the Cartesian coordinates represented by a rainstorm on September 14, 2008. The video images the RGB domain. To distinguish a river object from its were captured between noon (when the rain began) and background, histogram equalisation was used to enhance nightfall (when the flood tide occurred). Figures 4 and the object contrast. The MS algorithm was then used to 5 show the surveillance images of the Nan-Hu No. 2 divide the video images into buildings, plants, rivers, Bridge and Ma-Ling-Keng River used in the experiments, and embankments. Finally, the RegGro algorithm was respectively. used to segment the images to form a binary mask. The outlines of the river areas were preserved within the In this study, the proposed IFAM was used to binary mask. Figures 7 and 8 show example images from evaluate the test areas in various image sets to determine the two datasets obtained using the region-based image the optimal image dehazing and flood identification identification methods. methods. The automated identification method for conducting flood monitoring is based on real-time video The flood alarm system captures surveillance images images. The performance of the proposed single image from remote monitoring stations, transmitting the flood dehazing method was evaluated by applying it to recover monitoring information to an image server, and backing several hazy images. Figure 6 illustrates the experimental up the image data in an image database. The IFAM results of the proposed dehazing method, indicating continually calculates the image data and automatically that the proposed method can effectively remove haze, executes image dehazing procedures; it then executes the even in a region with high haze density. Therefore, the image analysis techniques to determine the water level of dark channel prior can be used to directly estimate the the current overflow range. Finally, the automatic flood- thicknessFigure 5:of Surveillancethe haze and consequentlyimage sequences obtain (froma high- top toalert bottom, monitoring left tosystem right) presents for the theMa-Ling-Keng detected results quality haze-free image. concerning the flooded areas. Hence, the proposed River dataset

(a) (b) (c) (a) (b) (c)

Figure 6: Image dehazing results: (a) original hazy images; (b) estimated transmission map and (c) restored haze-free images Figure 6: Image dehazing results: (a) original hazy images, (b) estimated transmission map, and (c)

Journalrestored of the haze-free National Science images Foundation of Sri Lanka 46(3) September 2018 336 Chien-Hao Tseng et al.

(a) (b) (c)

Figure 7: Surveillance image sequences of the Nan-Hu No. 2 Bridge. From left to right: (a) Input Figure 7: Surveillance image sequences of the Nan-Hu No. 2 Bridge. From left to right: (a) input haze-free images, (b) segmented images and (c) identified images haze-free images, (b) segmented images, and (c) identified images

September 2018 Journal of the National Science Foundation of Sri Lanka 46(3) Flood monitoring system using dehazing method 337

(a) (b) (c)

FigureFigure 8: 8:Surveillance Surveillance image image sequences sequences of the ofMa-Ling-Keng the Ma-Ling-Keng River. From leftR iver.to right: From (a) inputleft tohaze-free right: images, (a) Input (b) segmented images and (c) identified images haze-free images, (b) segmented images, and (c) identified images

Journal of the National Science Foundation of Sri Lanka 46(3) September 2018 338 Chien-Hao Tseng et al. system can provide real-time water resource information Conference on Networking, Sensing and Control, 23 – 25 from all monitoring stations to multiple users, thereby April, Ft. Lauderdale, FL, USA, pp. 757 – 761. supporting decision-related operations. DOI: https://doi.org/10.1109/ICNSC.2006.1673241 Chen S.P. (1990). Remote Sensing Analysis in Geoscience. Beijing Mapping Press, Beijing, China. CONCLUSION Chen Z., Shen J. & Roth P. (2013). Single image defogging algorithm based on dark channel priority. Journal of Image dehazing is a vital problem in monitoring systems. Multimedia 8(4): 432 – 438. This paper proposes an automated single image dehazing Fattal R. (2008). Single image dehazing. ACM Transactions on method based on the dark channel prior, in which haze Graphics 27(3): 1 – 9. is considered the transmission. An IFAM is proposed DOI: https://doi.org/10.1145/1360612.1360671 in which an automated single image dehazing method Gibson K.B., Võ D.T. & Nguyen T.Q. (2012). An investigation of dehazing effects on image and video coding. IEEE and an image analysis technology are integrated for Transactions on Image Processing 21(2): 662 – 673. flood monitoring. The module is designed to improve DOI: https://doi.org/10.1109/TIP.2011.2166968 the identification accuracy under adverse weather Guo F., Tang J. & Gai Z.X. (2014). Image dehazing based on conditions while preserving the accuracy under clear haziness analysis. International Journal of Automation and weather conditions. Haze and rain, which strongly Computing 11(1): 78 – 86. affect morphological features, are crucial features that DOI: https://doi.org/10.1007/s11633-014-0768-7 should be evaluated for achieving high identification He K., Sun J. & Tang X. (2010). Single image haze removal performance. The video surveillance images can be used using dark channel prior. IEEE Transactions on Pattern for implementing an unattended monitoring system for Analysis and Machine Intelligence 33(12): 2341 – 2353. detection and warning services. The experimental results Hughes D., Greenwood P., Blair G., Coulson G., Pappenberger demonstrate the suitability of the proposed automated F., Smith P. & Beven K. (2006). An intelligent and adaptable grid-based flood monitoring and warning system. IFAM, which can effectively detect and remove Proceedings of the UK eScience All Hands Meeting, haze artefacts from a single image and identify flood Nottingham, UK, pp. 53 – 60. hazard areas from image sequences. Kang L.W., Lin C.W. & Fu Y.H. (2012). Automatic single-image- based rain streaks removal via image decomposition. IEEE REFERENCES Transactions on Image Processing 21(4): 1742 – 1755. DOI: https://doi.org/10.1109/TIP.2011.2179057 Krzhizhanovskaya V.V. et al. (15 authors) (2011). Flood early Abeyratne V.D.K. & Ileperuma O.A. (2006). Air pollution warning system: design, implementation and computational monitoring in the city of Kandy: possible transboundary modules. Procedia Computer Science 4: 106 – 115. effects. Journal of the National Science Foundation of Sri DOI: https://doi.org/10.1016/j.procs.2011.04.012 Lanka 34(3): 137 – 141. Krzysztofowicz R. & Davis D.R. (1983). A methodology for DOI: https://doi.org/10.4038/jnsfsr.v34i3.3644 evaluation of flood forecast-response systems: 2. Theory. Albanese A., Disabato F., Vigna R. & Terzo O. (2008). Water Resources Research 19(6): 1431 – 1440. Early warning system for flood events and automatic DOI: https://doi.org/10.1029/WR019i006p01431 classification of low resolution satellite imagery: ITHACA Levin A., Lischinski D. & Weiss Y. (2008). A closed-form approach. Proceedings of the International Workshop on solution to natural image matting. IEEE Transactions Remote Sensing for Disaster Applications, Pavia, Italy, pp. on Pattern Analysis and Machine Intelligence 30(2): 1 – 6. 228 – 242. Attanayaka A. & Wijeyaratne S.C. (2013). Corticolous lichen DOI: https://doi.org/10.1109/TPAMI.2007.1177 diversity, a potential indicator for monitoring air pollution McCartney E.J. (1976). Optics of the Atmosphere: Scattering in tropics. Journal of the National Science Foundation of by Molecules and Particles. John Wiley and Sons, New Sri Lanka 41(2): 131 – 140. York, USA. DOI: https://doi.org/10.4038/jnsfsr.v41i2.5707 Namer E., Shwartz S. & Schechner Y.Y. (2009). Skyless Barnum P.C., Narasimhan S. & Kanade T. (2010). Analysis of polarimetric calibration and visibility enhancement. Optics rain and snow in frequency space. International Journal of Express 17(2): 472 – 493. Computer Vision 86(2 – 3): 256 – 274. DOI: https://doi.org/10.1364/OE.17.000472 Basha E.A., Ravela S. & Rus D. (2008). Model-based Narasimhan S.G. & Nayar S.K. (2003). Contrast restoration of monitoring for early warning flood detection, Proceedings weather degraded images. IEEE Transactions on Pattern of the 6th ACM Conference on Embedded Network Sensor Analysis and Machine Intelligence 25(6): 713 – 724. Systems, 5 – 7 November, Raleigh, USA, pp. 295 – 308. DOI: https://doi.org/10.1109/TPAMI.2003.1201821 DOI: https://doi.org/10.1145/1460412.1460442 Patel S.P. & Nakrani M. (2016). A review on methods of image Chang N. & Guo D.H. (2006). Urban flash flood monitoring, dehazing. International Journal of Computer Applications mapping and forecasting via a tailored sensor network 133(12): 975 – 8887. system. Proceedings of the 2006 IEEE International Rani M., Kumar P., Yadav M. & Hooda R.S. (2011). Wetland

September 2018 Journal of the National Science Foundation of Sri Lanka 46(3) Flood monitoring system using dehazing method 339

assessment and monitoring using image processing image defogging. Fujitsu Scientific and Technical Journal techniques: a case study of Ranchi, India. Journal of 50(1): 60 – 65. Geographic Information System 3(4): 345 – 350. Wang Y.F. (2009). Enhanced actions for water disaster warning DOI: https://doi.org/10.4236/jgis.2011.34032 system and evacuation. Proceedings of the Southeast Sapphaisal C. (2007). Forecasting and warning flood. Civil Asia International Program (SEAIP), 30 November - 04 Engineering Magazine, volume 2, pp. 38 - 49. Engineering December, Taichung, Taiwan, pp.1 ‒ 32. Institute of Thailand, Thailand. Wu J.H., Tseng C.H., Chen L.C., Lo S.W. & Lin F.P. (2015). Schechner Y.Y., Narasimhan S.G. & Nayar S.K. (2003). Automated image identification method for flood Polarization-based vision through haze. Applied Optics disaster monitoring in riverine environments: a case 42(3): 511 – 525. study in Taiwan. Proceedings of the AASRI International DOI: https://doi.org/10.1364/AO.42.000511 Conference on Industrial Electronics and Applications Selvi C. & Sathya S. (2014). Flood identification using satellite (IEA2015), London, UK, pp. 268 – 271. images. International Journal of Advanced Research in Xiao C. & Gan J. (2012). Fast image dehazing using guided Electrical, Electronics and Instrumentation Engineering joint bilateral filter. Visual Computer 28(6 – 8): 713 – 721. 3(1): 6497 – 6504. Zhai Y. & Ji D. (2015). Single image dehazing for visibility Shukla S. & Pandey G.N. (2014). To design an architectural improvement. Proceedings of the International Archives model for flood monitoring using wireless sensor network of the Photogrammetry, Remote Sensing and Spatial system. International Journal of Computer Science and Information Sciences, Toronto, Canada, pp. 355 – 360. Information Technologies 5(1): 502 – 507. Zhang J., Li L., Yang G., Zhang Y. & Sun J. (2010). Local Shwartz S., Namer E. & Schechner Y.Y. (2006). Blind haze albedo-insensitive single image dehazing. Visual Computer separation. Proceedings of the IEEE Conference on 26(6 – 8): 761 – 768. Computer Vision and Pattern Recognition, 17 – 22 July, Zhang J., Zhou C., Xu K. & Watanabe M. (2002). Flood New York, USA, pp. 1984 – 1991. disaster monitoring and evaluation in China. Environmental DOI: https://doi.org/10.1109/CVPR.2006.71 Hazards 4(2): 33 – 43. Tan Z., Bai X., Wang B. & Higashi A. (2014). Fast single- DOI: https://doi.org/10.3763/ehaz.2002.0404

Journal of the National Science Foundation of Sri Lanka 46(3) September 2018