A Maritime Cloud-Detection Method Using Visible and Near-Infrared Bands over the Yellow Sea and Bohai Sea

: Accurate cloud-masking procedures to distinguish cloud-free pixels from cloudy pixels are essential for optical satellite remote sensing. Many studies on satellite-based cloud-detection have been performed using the spectral characteristics of clouds in terms of reﬂectance and temperature. This study proposes a cloud-detection method using reﬂectance in four bands: 0.56 µ m, 0.86 µ m, 1.38 µ m, and 1.61 µ m. Methodologically, we present a conversion relationship between the normalized difference water index (NDWI) and the green band in the visible spectrum for thick cloud detection using moderate-resolution imaging spectroradiometer (MODIS) observations. NDWI consists of reﬂectance at the 0.56 and 0.86 µ m bands. For thin cloud detection, the 1.38 and 1.61 µ m bands were applied with empirically determined threshold values. Case study analyses for the four seasons from 2000 to 2019 were performed for the sea surface area of the Yellow Sea and Bohai Sea. In the case studies, the comparison of the proposed cloud-detection method with the MODIS cloud mask (CM) and Cloud-Aerosol Lidar and Infrared Pathﬁnder Satellite Observation data indicated a probability of detection of 0.933, a false-alarm ratio of 0.086, and a Heidke Skill Score of 0.753. Our method demonstrated an additional important beneﬁt in distinguishing clouds from sea ice or yellow dust, compared to the MODIS CM products, which usually misidentify the latter as clouds. Consequently, our cloud-detection method could be applied to a variety of low-orbit and geostationary satellites with 0.56, 0.86, 1.38, and 1.61 µ m bands.


Introduction
Cloud information is significant for weather analysis, climate change, and other shortterm environmental issues [1,2]. Satellite remote-sensing plays a crucial role in monitoring the Earth's surface and atmosphere. For remote sensing of the Earth's surface, clouds pose an obstacle, as they cover a large area of the Earth's surface and need to be effectively removed. Thus, cloud detection is a preliminary prerequisite step and one of the most difficult problems in satellite remote-sensing [3,4].
Many studies on satellite-based cloud detection have been performed which measured the spectral characteristics of clouds, such as high reflectance and low temperature in the visible (VIS), near-infrared (NIR) [5], shortwave infrared (SWIR) [6], and longwave infrared (LWIR) [7][8][9] bands. Brightness temperature difference (BTD) methods are also used for cloud detection in satellite remote sensing. For example, a BTD between 11 and 12 µm is useful for detecting stratiform clouds such as cirrus [10]. The BTD between 3.9 µm and 11 µm is useful for detecting low-level water clouds in most scenes and partial or thin cloud coverage at night [10]. Furthermore, there have been a number of studies to accurately identify clouds over ice and water using the SWIR band [11], mid-wave infrared (IR) bands [12], VIS bands, and NIR bands [13]. Recently, cloud detection has been explored through various deep learning techniques, including convolutional neural networks [14,15], machine learning methods, such as the support vector machine [16] and random forest method [17], as well as classical cloud detection algorithms including threshold-based technique [18] and dynamic threshold [19].
Traditionally, cloud-detection methods based on simple VIS and IR window threshold values have been widely applied [20,21] in many projects, such as the International Satellite Cloud Climatology Project, Clouds from the Advanced Very High Resolution Radiometer (AVHRR; National Oceanic and Atmospheric Administration), and AVHRR Processing scheme Over cLouds, Land and Ocean [22], because clouds are generally characterized by a higher reflectance and lower temperature than the underlying Earth's surface [23].
In addition, NASA's earth-observing system anchored several active and passive satellite sensors with unprecedented observation capabilities in a formation called the afternoon constellation or the A-Train [24]. The A-Train has been regarded as among the most accurate sensors for detecting clouds globally [25]. The Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) also provides global cloud and aerosol data with the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) onboard CALIPSO, observing the backscattered signal at 0.532 and 1.064 µm bands [26]. The Moderate Resolution Imaging Spectroradiometer (MODIS) cloud mask (CM) product has been the most fundamental of the MODIS land, ocean, and atmosphere products [27,28]. This is because of its long temporal record, wide swath and sampling, and broad spectral range [25]. The MODIS CM was established based on global analyses using calibrated MODIS radiance data in the VIS and IR bands [28].
The MODIS products provide the cloud detection product referred to as MOD35 [29,30], relying on a series of sequential tests on reflectance (R) and radiances (or brightness temperatures (BT)) in the instrument field of view (FOV). The MOD35 CM includes many cloud test thresholds of reflectance, such as R 0.67µm and R 0.87µm for thick low and high clouds, R 1.38µm for high thin and thick clouds, and R 1.6µm for low and high clouds over snow. Additionally, the MOD35 CM includes many IR cloud test thresholds; such as BT 11µm for oceans, BT 13.9µm for all nonpolar regions, BT difference (BTD) BT 6.7µm for all polar regions except Antarctic night, surface temperatures at BT 11µm , BT 11µm−12µm for all land except Antarctica, BT 8.6µm−11µm for all oceans, BT 11µm−3.9µm for day snow, ice, land, and ocean, BT 3.9µm−12µm for day snow, ice, and land, BT 7.2µm−11µm for night ice, land, and ocean, and BT 8.6µm−7.2µm for night ocean [31]. The thresholds of the MODIS cloud detection algorithm vary depending on the region or conditions [31][32][33]. Consequently, the MOD35 CM products provide a variety of information on cloud mask, sun glint, snow/ice, heavy aerosol, thin cirrus, shadow, and information on whether it is day/night and if the clouds are present over land and water [30]. However, the problem with the MODIS CM is that it misidentifies snow [30,32,34] and ice [35] as clouds.
The objective of this study is to explore a method for detecting daytime clouds over the ocean. We selected the Yellow Sea and Bohai Sea as the study areas. We used only four channels of MODIS sensors in the VIS and NIR spectral bands. We verified our method for several case studies by comparing it with MODIS CM. Figure 1 shows the MODIS red-green-blue (RGB) image using 0.56, 0.86, and 1.6 µm bands [36] and the MODIS CM (MOD35_L2) over the Yellow Sea and Bohai Sea on 3 January 2013, 03:15 UTC. The current MODIS CM misidentifies sea ice over the Bohai Sea as clouds (pink).

Cloud Detection Method
This study used four bands-one VIS (0.56 µm, green band) and three NIR (0.86, 1.38, and 1.61 µm) bands for cloud detection. For thick cloud detection, we used the normalized difference water index (NDWI), a satellite-derived index from the combination of the green band and NIR band [37,38]. The green and NIR bands of the Landsat Thematic Mapper were used to produce the NDWI [39]. The water surface reflects the most radiance in VIS bands but absorbs strongly in NIR bands [38]; thus, NDWI is sensitive to liquid water changes.
This study proposes a relationship between NDWI and the 0.56 µm band for cloud detection over the sea surface. Our thick cloud detection method consists of the following steps: First, we chose cloud pixels from the MODIS RGB image. Figure 2 presents an example of Terra MODIS RGB images on 1 June 2017, 02:15 UTC used in this study. Second, we adopted the NDWI. In this study, NDW I obs indicated the NDWI values calculated using the green and NIR bands of MODIS data as follows [37]: Here, NDW I obs is the NDWI observed from the MODIS reflectance data. R 0.56µm and R 0.86µm are the reflectances of 0.56 and 0.86 µm, respectively.
Third, we presented NDW I cal , which is a regression relationship fitting the distribution of the observed pixels in the NDWI-R 0.56µm plane. Herein, NDW I cal was defined in a quadratic form as follows: (2) where NDW I cal is the NDWI calculated from reflectance data of the MODIS 0.56 µm band in the NDWI-R 0.56µm plane. a 0 , a 1 , and a 2 are regression coefficients.
Real cloud pixels were distributed around the NDW I cal curve (Equation (2)) (see Figure 5). Thus, we determined the threshold values of the NDWI for cloud detection using the standard deviation (σ) of the Gaussian distribution of cloud pixels around the NDW I cal curve (Equation (2)), for the area marked by the pink box (in Figure 2), as follows: where k is the coefficient empirically determined from the histogram distribution of the pixels for clouds. In this study, we set k = 1. σ 1 is the standard deviation of the histogram distribution of pixels in the cloud. Fourth, the method based on Equations (1)-(3) without other band observations often misidentified thin clouds as non-cloud pixels. Thus, we combined the NDWI-based method for thick cloud detection with the threshold methods using 1.38 and 1.61 µm bands for thin cloud detection. The 1.38 µm band is commonly called the "cirrus band" because of its sensitivity to water vapor above and within the cirrus [40,41] and is therefore useful for optically thin cirrus detection [42][43][44][45][46]. The 1.61 µm band has a lower reflection and an absorption effect on cloud ice particles [41]. This 1.61 µm band is often used in the form of Normalized Difference Snow Index [37][38][39].
We applied thresholds to the reflectances R 1.38µm and R 1.61µm at 1.38 and 1.61 µm bands, respectively, as: where σ 2 and σ 3 are the threshold values empirically determined from the histogram distributions of the pixels for thin clouds at the 1.38 and 1.61 µm bands, respectively. For this study, we used σ 2 = 0.006 and σ 3 = 0.04, which were chosen from statistical comparison results (the best values of correlation coefficient (CC) and root mean square error (RMSE)) between MODIS CM and our CM including 1.38 and 1.61 µm bands.

Comparison
The proposed cloud detection method is compared with data from MOD35_L2. MODIS detected various types of clouds, including thin cirrus clouds. The statistical analysis of the proposed cloud detection algorithm with MODIS35_L2 was performed using the probability of detection (POD), false-alarm ratio (FAR), and Heidke Skill Score (HSS) as follows: where A indicates that both the proposed CM and MODIS CM indicate a cloud pixel; B implies that the proposed CM is a cloud pixel, but MODIS CM is not a cloud pixel; C infers that the proposed CM is not a cloud pixel, but MODIS CM is a cloud pixel; and D infers that both CMs are not clouds. The POD and HSS are close to 1, and the FAR is close to 0 for the best possible and perfect detections [47]. The contingency tables are presented in Table 1.  Figure 3 shows a flowchart of the cloud detection method presented in this study. First, the MODIS calibrated radiance data (MOD021KM) at the VIS and NIR bands were converted to reflectance values. Next, the cloud and no-cloud flags were retrieved using the proposed cloud detection algorithm, using the NDWI, the green band, 1.38 µm, and 1.61 µm band reflectances. Finally, our CM method was compared with MODIS CM.

MODIS
In this study, we used data from the MODIS onboard Terra satellite to develop a cloud detection method. The MODIS sensor has 36 spectral bands from VIS (0.4 µm) to IR (14 µm). This study selected four MODIS bands; 0.56 µm (band 4), 0.86 µm (band 2), 1.64 µm (band 6), and 1.38 µm (band 26). MODIS CM (MOD35_L2) was used to compare the results of the cloud detection results proposed in this study. This study used "cloudy" and "probably cloudy" as a classification of cloudiness in MODIS CM based on the unobstructed FOV quality flags ("cloudy," "probably cloudy," "probably clear," and "confident clear") from the MODIS CM.
We chose the Yellow Sea and Bohai Sea regions as the study area for the cloud detection method development because sea ice and clouds often coexist in this region during winter. The ice crystals of clouds have properties similar to those of ice over the sea and hence are difficult to distinguish, making this region suitable for testing the cloud detection algorithm. Figure 4 displays the study area (in the red box) selected for cloud detection in the Yellow Sea and the Bohai Sea. Table 2

CALIPSO
This study used the CALIPSO vertical feature mask (VFM) data (version 4.200 of CAL_LID_L2_01km CLay-Standard data) with cloud layer information at a 1 km spatial resolution to validate the proposed CM because cloudy pixels are included in the CALIPSO data feature classification flags. To ensure spatial and temporal colocation, CALIPSO laser footprints that were found within the bounds of and acquired at nearly the same time as the MODIS image were chosen.

Results
Figure 5a displays the distribution of cloud pixels (pink color) selected in Figure 2 in the NDWI-green band plane. Cloud pixels are distributed within the range of NDWI ≤ 0 and green band reflectance > 0.3 in the NDWI-green band plane. Figure 5b shows the relationship between NDWI and the reflectance of the green band (0.56 µm) for cloud detection. The blue pixels indicate cloud pixels, including thin and thick clouds. The red line indicates the regression relationship obtained using (2). In this case, the regression coefficients a 0 , a 1 , and a 2 were calculated as 0.079, −0.4, and 0.312, respectively. The standard deviation value (σ), including the uncertainty of cloud detection using the relationship between NDWI and MODIS R 0.56µm , was estimated to be 0.0377 using Equation (3)     We applied the regression coefficients (a 0 = 0.079, a 1 = −0.4, and a 2 = 0.312) and σ = 0.0377, which were obtained from MODIS data on 1 June 2017, 02:15 UTC for all the cases. The first column shows the MODIS RGB images, and the second column displays the results of cloud detection when we applied only the NDWI-green band relationship to our CM algorithm. We can identify that the NDWI-green band relationship was useful for detecting thick clouds but missed parts of thin clouds. The third column shows the results using R 1.38µm , which detected cirrus clouds. The fourth column displays the results using R 1.61µm , which detected most of the thin and high clouds. The fifth column shows the final results of our CM including the NDWI-green band relationship, R 1.38µm , and R 1.61µm , which detects thin and thick clouds. Consequently, the NDWI-green band relationship in the proposed cloud-detection method detected fewer cloud pixels than the MODIS CM in the Yellow Sea and Bohai Sea because the proposed method excludes semitransparent clouds and optically thick clouds. However, the proposed method based on the combination of NDWI-green band relationship and threshold values of 1.38 and 1.61 µm bands was much better at detecting thin clouds than the MODIS CM, as shown in Figure 7. These results demonstrate that the proposed CM requires 0.56 µm, 0.86 µm, 1.38 µm, and 1.61 µm bands for the best detection of clouds.       Figure 8. We classified the vertical atmospheric and surface features of CALIPSO observations based on the feature classification flags of the CALIPSO VFM data. In the legend for Figure 9, no signal, subsurf, surface, strato, aerosol, cloud, and clear indicate no signal (complete signal attenuation), subsurface, surface, stratospheric aerosol, tropospheric aerosol, cloud, and clear air, respectively. The MODIS CM and the proposed CM were in qualitatively good agreement with the CALIPSO VFM data for the four cases. The statistical scores of POD, FAR, and HSS also show quantitatively consistent results between the MODIS CM and the proposed CM. Table 4 summarizes the validation results using the CALIPSO data in Figure 8, which demonstrates that the proposed CM is consistent with the MODIS CM.

Discussion
The proposed method has the advantage of being able to distinguish sea ice from clouds. Figure 12   The proposed method has the additional advantage of being able to distinguish dust from clouds. Figure 13

Conclusions
Cloud information plays an important role in weather analysis, climate change, and many other geo-science applications. Currently, satellite remote-sensing is widely used to detect and analyze cloud information. An accurate cloud-masking procedure to distinguish cloud-free pixels from cloud pixels is essential for the unhindered analysis of many parameters on the Earth's surface, such as ocean color study in satellite remote-sensing. Generally, cloud-detection algorithms depend on the spectral characteristics of clouds, such as high reflectance and low temperature in the VIS, NIR, SWIR, and LWIR bands. Generally, split-window measurements, in other words, BTD in IR bands such as 10.8 µm and 12.0 µm depending on thresholds, have been widely used to detect or mask clouds from satellite observations. In particular, MODIS CM has been widely used because of its long temporal record, wide swath and sampling, and broad spectral range. However, it is well known that the MODIS cloud detection algorithm often misidentifies clouds when sea ice is present.
Thus, this study proposes a unique cloud-detection method using the relationship between NDWI and reflectance of the green band with the uncertainty of standard deviation in addition to 1.38 and 1.61 µm bands for thin cloud detection. Multiple case studies using MODIS observation data were performed for the spring, summer, autumn, and winter seasons from 2000 to 2019 over the Yellow Sea and Bohai Sea. The comparison of the proposed cloud detection method with MODIS CM showed improved performance and detailed cloud information compared to the MODIS CM. For the period 2000-2019, the average POD, FAR, and HSS were 0.853, 0.069, and 0.719, respectively. Relatively poor statistical results often appear due to the coexistence of clouds and sea ice in the winter and haze or dust in the spring. The time series analysis using the 79 case studies from 2000 to 2019 demonstrated that the POD and HSS values tended to increase, while the FAR values decreased between MODIS CM and our CM. A few cases displayed relatively large discrepancies due to the presence of haze or semitransparent clouds in the scene and seasonally turbid water in the study area. In addition, the proposed cloud-detection method has the advantage of distinguishing cloud pixels from sea ice or yellow dust, which is a difficult problem that the current MODIS CM algorithm cannot mitigate.
Visual inspection of MODIS data indicated the presence of atmospheric cover resembling haze or semitransparent clouds at the location of most case studies. The classification of these pixels as clouds often depends on the user's need for their own atmospheric or sea-surface studies. Currently, the thresholds used in our method for thin clouds have not been optimized. Thus, the proposed approach occasionally identified optically thin clouds, despite visual inspection determining them to be cloud-free. However, these disadvantages may be overcome with additional modification of thresholds and the use of other spectral bands sensitive to various types of clouds according to user needs.
Consequently, the proposed cloud detection method can be applied to a variety of other optical sensors with VIS and NIR bands, such as the Geostationary Ocean Color Imager-II from 0. 38