A thin cloud removal method from remote sensing image for water body identification

In this paper, a thin cloud removal method was put forward based on the linear relationships between the thin cloud reflectance in the channels from 0.4 μm to 1.0 μm and 1.38 μm. Channels of 0.66 μm, 0.86 μm and 1.38 μm were chosen to extract the water body information under the thin cloud. Two study cases were selected to validate the thin cloud removal method. One case was applied with the Earth Observation System Moderate Resolution Imaging Spectroradiometer (EOS/MODIS) data, and the other with the Medium Resolution Spectral Imager (MERSI) and Visible and Infrared Radiometer (VIRR) data from Fengyun-3A (FY-3A). The test results showed that thin cloud removal method did not change the reflectivity of the ground surface under the clear sky. To the area contaminated by the thin cloud, the reflectance decreased to be closer to the reference reflectance under the clear sky after the thin cloud removal. The spatial distribution of the water body area could not be extracted before the thin cloud removal, while water information could be easily identified by using proper near infrared channel threshold after removing the thin cloud. The thin cloud removal method could improve the image quality and water body extraction precision effectively.


Introduction
Water body is a significant parameter for earth surface environment, climate change and flood disaster monitoring. Satellite remote sensing technology can play an important role in water body monitoring and assessment because of its ability of quickly acquiring wide range of earth surface information, including areas being hard to reach (Zheng, 2008). Optical remote sensing data, such as Moderate Resolution Imaging Spectroradiometer (MODIS) and Advanced Very High Resolution Radiometer (AVHRR), have had many successful application cases for water body monitoring (Barton and Bathols, 1989;Sheng et al., 1998;Sheng et al., 2001;Bryant and Rainey, 2002;Zhan et al., 2002;Peng et al., 2004;Cai et al., 2005;Jain et al., 2006;Huang et al., 2008). But for the optical remote sensing data, only cloud-free images are required to monitor the water under a clear sky. However, there is little clear sky during the flood, and cloud contamination will result in invalidation of optical remote sensing data. If the water body information under the cloud can be extracted, it will be a great significant thing. For optical remote sensing data, when affected by the thick clouds, the ground information can not be got only using the image itself; when affected by thin cloud, part of the energy reflected by surface can get through the thin cloud and reach the sensor. In this case, the surface object, especially the water body, can be identified after removing the impact of thin cloud.
Many removal methods for thin cloud have been studied so far. Tasseled-cap transformation was an orthogonal transformation of the original data into new multi-dimensional space. To the Landsat/MSS or TM images contaminated by the cloud, the fourth tasseledcap band was mainly taken as the noise (cloud). Then, the first three tasseled-cap bands were transformed back to RGB image which had removed the impact of cloud (Richter, 1996;Ma and Gu, 2005). But this method would lead to the loss of band information and be unfit to the image with some bright subjects. The algorithm based on wavelet theory was used to remove the cloud (Tseng et al., 2008;Zhu et al., 2009). The algorithm could make earth surface information and cloud noises distributed to the coefficients of low levels and high levels respectively. Most of cloud noises and some useful information were included in high level coefficients where cloud was brighter than the surface information. Cloud could be removed from these coefficients and information. Weight factors were assigned to the detailed coefficients of low level, high level, and approximation coefficients respectively for enhancing the contrast of earth surface and decreasing remaining cloud. The three parts of coefficients were reconstructed and fused to get processed results. The homomorphism filtering method had also been used to remove the thin cloud because the thin cloud was the low frequency component in remote sensing images (Zhao and Zhu, 1996). However, the above methods all changed the physical information of the original image, which made the quantitative analysis difficult. An improved homomorphism filtering method was adopted to remove thin cloud qualitatively at first, and all bands were selected to classify pixels (Feng et al., 2004;Ma and Gu, 2005). Based on the fact that visible band images contained more aerosol effects than infrared bands, visible bands were used to determine clear or hazy regions and the reflectance of different surface classes in clear regions were calculated. At last the mean reflectance of the same classes in clear and hazy regions was matched to remove cloud. Atmospheric correction methods could remove the effects of aerosol, cloud and water vapor (Liang et al., 2001;Liang et al., 2002;Juan et al., 2010), and met the requirements of quantitative analysis to some extent, while its computation was too complicated to execute. Based on the above analysis, we can know that there is not a proper thin cloud removal method for easily and quickly extracting water information from the remote sensing image contaminated by the thin cloud. In this paper, we put forward an effective and easy thin cloud removal method according to the spectral characteristics of water body, atmosphere and thin cloud, and the relationships between the thin cloud reflectance of the channels from 0.4 µm to 1.0 µm and 1.38 µm. The thin cloud removal method can be used to improve the image quality and water body extraction precision for flood disaster monitoring and water resource management.

Study area
In this paper, for researching the effectiveness of the thin removal method, two study areas were chosen. The first study area is within the Huaihe River Basin (Fig. 1), with 800 km in length and 400 km in width, covering 270 000 km 2 . Huaihe River is the sixth largest river in China, located in the middle eastern part of China. In recent 20 years, three severe floods have happened respectively in 1991, 2003 and 2007. The disaster has made many adverse impacts on economic development in this region. For the first study area, Earth Observation System Moderate Resolution Imaging Spectroradiometer (EOS/MODIS) data was applied.
The second study area was situated within the Danjiangkou Reservoir (Fig. 1). Danjiangkou Reservoir is a multi-purpose reservoir, located in Xichuan County of Henan Province and Danjiangkou City of Hubei Province, central China. It was constructed in 1958, created by the Danjiangkou Dam, and at the time was one of the largest reservoirs in Asia. The Danjiangkou Reservoir is the water source for one of the routes of the ′South to North Water Transfer Project′, providing water to Beijing and Tianjin. Moreover, it serves as a water supply for the local region, as well as irrigation, electricity generation and flood control. For this study area, Fengyun-3A (FY-3A) Medium Resolution Spectral Imager (MERSI) and Visible and Infrared Radiometer (VIRR) data were used.
It is quite significant to monitor the two study areas by using satellite remote sensing data for flood disaster prevention and water resource management.

Data and processing 2.2.1 MODIS data and processing
The MODIS instrument has a viewing swath width of 2330 km and scans the entire surface of the Earth every one to two days. Its detectors measure 36 spectral bands between 0.405 µm and 14.385 µm, including two channels of 0.66 µm and 0.86 µm with 250 m resolution, five channels with 500 m resolution and 29 channels with 1 km resolution (Justice et al., 1998;Liu and Ge, 2000;Liu and Yang, 2004). MODIS data with the spatial resolution from hundred meters up to kilometer, and the temporal resolution of two or more days, have the irreplaceable ability of quickly assessing and mapping surface water body. In this paper, 0.66 µm, 0.86 µm and 1.38 µm channels will be applied to identifying the water body under the thin cloud. In order to match the 250 m spatial resolution of the 0.66 µm and 0.86 µm channels, 1.38 µm channel with 1 km resolution was resampled to 250 m. The MODIS images contaminated by the thin cloud on June 25, 2003 in the Huaihe River Basin were selected as the test data.

FY-3A data and processing
The development of China′s new generation of polar-orbiting meteorological satellite series FY-3A greatly enhances the surface monitoring ability, meanwhile provides a new data source for water body monitoring for large area. Medium Resolution Spectral Imager (MERSI) as one of the 11 instruments on board the FY-3A satellite has five channels with 250 m spatial resolution (including 0.66 μm and 0.86 μm channels) and 15 channels with 1 km resolution. MERSI has both multi-spectral and high spatial resolution properties, which greatly improves the ability to observe fine surface features Yang et al., 2009). Therefore, MERSI instrument has greater potential to provide the ideal data sources for flood and water resource monitoring. Visible and Infrared Radiometer (VIRR) on board the FY-3A satellite has the 1.38 μm channel with 1 km resolution. VIRR and MERSI have the same scene time and area, so VIRR can detect the cloud information when MERIS data is used to map water body. For matching the spatial resolution of the MERSI data, 1.38 µm channel of VIRR was resampled to 250 m resolution. The FY-3A images contaminated by the thin cloud for the Danjiangkou Reservoir on May 4, 2011 were chosen as the second test data.

Cloud detection
Cloud detection is the first step to remove the influence of cloud. Confirmation of thin cloud region can avoid the correction of the free-cloud area. Thin cloud consists of ice particles having different sizes and shapes. The ′effective′ particle sizes (radii of equivalent spheres) are usually greater than 5 μm. The ice particles in the spectral range of 0.4-1.0 μm are much larger than wavelengths and ice is non-absorbing. Weak ice absorptions occur near 1.24 μm and 1.38 μm. The 1.38 μm channel has strong capabilities of thin cloud detection (Ackerman et al., 1998;Liu and Yang, 2004). The reflectances at 1.38 μm are smaller than reflectances in the 0.4-1.0 μm spectral mainly because of absorption by water vapor above and within the cirrus clouds (Gao et al., 1998). Water vapor almost abstracts all the spectral energy, most of reflectance come from the high-altitude thin cirrus in the 1.38 μm channel. The thin cloud information can be extracted easily through setting threshold value in the 1.38 μm channel (Ackerman et al., 1998), so the threshold method was employed to distinguish the thin cloud in this paper (Equation (1)).
where T L_1.38 is the threshold value between cloud and clear sky in the 1.38 μm channel; R 1.38 is the reflectance in the 1.38 μm channel; T M_1.38 is the threshold value between thin cloud and thick cloud in the 1.38 μm channel. However, 1.38 μm is not sensitive to all types of cloud, while 0.66 μm can be utilized to detect the thick cloud which have higher reflectance than other surface objects in the visible channel (Chen et al., 2002), so the threshold of 0.66 μm will be added to distinguish the thick cloud from other objects, the expression is as follows.
where R 0.66 is the reflectance in the 0.66 μm channel; T 0.66 is the threshold value between thick cloud and clear sky in the 0.66 μm channel.

Thin cloud removal
Hypothesizing that atmosphere in the study area is homogeneous, we think that the effect of atmosphere is much smaller than thin cloud for water extraction. So, surface reflection and atmospheric scattering to the sensor are considered together, expressed as L s + a (Zheng et al., 2007). Thus, radiation brightness (L * ) received by the sensor can be considered primarily including two components (Fig. 2), cloud reflection (L c ), and surface reflection and atmospheric scattering (L s + a ), then L * can be expressed as: For a flat and Lambertian surface, L * can be expressed as: where R * is the apparent reflectance; E is the exoatmospheric solar irradiance; θ is the sun zenith angle. L c can be calculated by the following equation: where R c is the reflectance of thin cloud. L s + a can be expressed as: where T c is the thin cloud transmittance; R s + a is the reflectance of surface and atmosphere. From equations (3) to (6), we can obtain the following expression.
The previous researchers found that there was a good linear relationships between the thin cloud reflectance of the channels from 0.4 μm to 1.0 μm and the 1.375 μm channel of AVIRIS data (Gao et al., 1998;Gao and Li, 2000). The reflectance of thin cloud can be expressed as: where R c(λ) is the reflectance of thin cloud in band length λ; R c(1.38) is the reflectance of thin cloud in the 1.38 μm; k is the empirical coefficient which can be got by the reflectance relationships of 1.38 μm, visible and near infrared channels (Gao et al., 1998); T c reflects the transmittance of thin cloud, which can be calculated by: where n is an experiential parameter with a typical value between 1 and 2, which depends on both the solar and sensor geometries.
To remote sensing data, it is more difficult to distinguish water body from other surface objects in the visible channel than in the near infrared channel. Using the threshold value method, water body information can be extracted easily in the near infrared channel. Because 0.86 µm is sensitive to the water body, and has a good relationship with the 1.38 µm channel, therefore, 0.86 µm was chosen as the main channel to extract water body information in this paper． Figure 3 showed that the west part of MODIS image on June 25, 2003, was contaminated severely by cloud. Firstly, we used 1.38 μm channel to detect cloud. How-ever, 1.38 μm channel was not sensitive to all kinds of clouds. Thick cloud at the middle-east part of image in 0.66 μm could be seen obviously (Fig. 3a), but almost invisible in 1.38 μm (Fig. 3b). So the cloud undetected by 1.38 μm was identified by 0.66 μm (channel 1) effectively based on the Equation (2).

Thin cloud removal from MODIS image
For accurately detecting the cloud covered area, the selection of threshold value is important. Referring to the spectra curve of thick and thin clouds (Gao et al., 1998;Gao and Li, 2000), for 1.38 μm channel, 0.01 was selected as the threshold value to distinguish cloud and cloud-free areas, and 0.13 was selected as the threshold value to distinguish thin cloud and thick cloud. For 0.66 µm channel, 0.30 was selected as the threshold value to detect the thick cloud. Figure 4 showed the scatter map of images in the 0.86 µm and 1.38 µm channels. The X axis mainly represented the surface reflectance, and Y axis indicated the cloud reflectance. The left scatter points reflected the relationship between lower surface reflectance and thin cloud was chosen to estimate k value, which was set as 0.7. It is noted that if selected k value is higher, then the corrected surface reflectance will be lower, and if selected k value was smaller, the correction of thin cloud was not enough. By comparing the same surface type reflectance under clear sky and thin cloud, T c could be estimated, and then n was estimated to be equal to 1 according to several experiments based on the Equation (9). Figure 5 showed the images comparison results before and after the thin cloud removal. It was very obvious that the image after the thin cloud removal became clearer, which indicated that the thin cloud removal   Fig. 6. Through profile analysis, we knew that the algorithm did not change the surface reflectance under clear sky. To the area contaminated by the thin cloud, the reflectance decreased after the thin cloud removal, which proved the efficiency of thin cloud removal method.
In order to verify the impact of classification before and after the thin cloud removal, a sub-area image in the Huaihe River Basin contaminated by the thin cloud in the 0.86 μm was chosen as the test area shown in Fig. 7. Threshold method was employed to classify the two images before and after the thin cloud removal (Fig. 7a,  7b). In the classification process, we found that the threshold method was invalid and could not extract the exact spatial distribution of the water body area before the thin cloud removal; while water information could easily be identified after the thin cloud removal by using threshold method (Fig. 7c). An image under the clear sky on June 28, 2003 was selected to extract water body  Comparison of surface reflectance before and after the thin cloud removal can reflect the effect of thin cloud removal. The clear sky image on June 28, 2003 was selected as the reference reflectance to compare with the water reflectance before and after the thin cloud removal, as shown in Table 1. The reflectance before the thin cloud removal was significantly higher than the reference value under the clear sky. After the thin cloud removal, the reflectance value of water was closer to the reference reflectance.

Thin cloud removal from MERSI and VIRR images
Based on the thin cloud removal method, we used MERSI and VIRR images to extract the water body information in the Danjiangkou Reservoir on May 4, 2011.
Detailed process of the thin cloud removal was similar to that for MODIS images. The parameter k was estimated to be 0.4, and n was equal to 1. Compared with the image under the clear sky on May 4, 2011, the image after the thin cloud removal became clearer in the south-west region of the reservoir (Fig. 8), which indicated that image quality was improved. Figure 8 obviously showed that quality of image was better after the thin cloud removal, and water information could be easily extracted by threshold method, which indicated the thin cloud removal method was reliable for FY-3 data. For quantitatively describing the effect of thin cloud removal to FY-3 image, three representative samples were chosen ( Table 2). The image under clear sky image on May 5, 2011 was selected as the reference reflectivity, to compare with the surface reflectance before and after the thin cloud removal, as shown in Table 2. The reflectance before the thin cloud removal was significantly higher than reference reflectivity under the clear sky. After the thin cloud removal, values of the first and third sampling points values were closer to the reference reflectance. The second sampling point did not changed because no thin cloud contamination. This point value was much higher than reference reflectance under the clear sky because of atmospheric scattering. The comparison result indicated effectiveness of the thin cloud removal method. Note: locations of sampling points 1, 2 and 3 were showed in Fig. 8

Discussion
Both MODIS and MERSI data, with the spatial resolution from hundred meters up to kilometers, and the temporal resolution of two or more days, have the irreplaceable ability of quickly assessing and mapping the flood and water. Under the thick cloud, the ground information can not be got by using the optical remote sensing data, while multi-sensor data involving optical and microwave data can resolve this problem effectively. When the optical remote sensing data are affected by the thin cloud, part of the energy reflected by surface can get through it and reach the sensor, and then ground information can be got by using the thin cloud removal method. Atmospheric correction could remove the effects of aerosol, cloud and water vapor. But the computation procedures was quite complicated. In order to easily and quickly extract water information, although we possibly get better result after atmospheric correction, the effect of atmosphere is much smaller than thin cloud for water extraction. So, supposed that atmosphere in the study area is homogeneous, surface reflection and atmospheric scattering to the sensor are both considered. This ensured that we focused on the thin cloud removal and did not attempt to correct the atmospheric effect.
Cloud detection is the first step for removing the effect of cloud. Water vapor almost abstracts all the spectral energy, most of reflectance come from the high-altitude thin cirrus in the 1.38 μm channel. The 1.38 μm has been used to detect the high-altitude thin cirrus. The 0.66 μm channel was utilized to distinguish the thick cloud with higher reflectance than other surface objects.
To determine the threshold value, the spectra curve of thick and thin clouds were referred. Application of more channels such as near 3.7 μm and 12.0 μm may improve the cloud detection precision. The study about cirrus ice particle size, shape, and orientation distributions, spatial variability, and scattering phase functions were not sufficient for reliable modeling of cirrus reflectivity. Cirrus effects were difficult to simulate by using radiative transfer model. So, the empirical relationships were selected to characterize the main cirrus properties. In the thin cloud removal method, k and n were empirical coefficients, different data have different k and n values, which possibly brought the errors into the thin cloud removal model, so the calibration of the two coefficients needs numerous experiments. It will be our main efforts to quantify the parameters in the further work.

Conclusions
The simple thin cloud removal method was proposed based on the linear relationships between the thin cloud reflectance of the channels from 0.4 µm to 1.0 µm and 1.38 µm. In order to validate the effectiveness of this method for different areas and data, two application examples were given, one was for EOS/MODIS data and the other was for FY-3A/MERSI and VIRR data. According to the results of extracting the water body information, it can be concluded that the impact of cloud has been weakened after the thin cloud removal, whist the data of the area without the thin cloud did not be changed. The thin cloud removal algorithm could meet the requirements of quantitative analysis for further analyses. The image quality after the thin cloud removal was obviously better than that before the thin cloud removal. Before the thin cloud removal, the threshold method was invalid and could not extract the exact spatial distribution of the water body area when image classification; while water information could easily be identified by using threshold method after the thin cloud removal. The application results indicated that thin cloud removal model was effective to improve the quality of EOS/MODIS, FY-3A/MERSI and VIRR images as well as the water body extraction precision, which promoted the application of China′s polar-orbiting meteorological satellite FY-3.