A method for the spectral analysis and identification of Fog , Haze and Dust storm using MODIS data

The three typical extreme weather of fog, haze and dust storm have occurred frequently in recent years in 10 China. These events influence the transportation, the ecological environment, and the daily lives of people. Remote sensing is very important technology that can be used to monitor them due to its high temporal resolution and wide area of coverage. But because the spectral features of the three extreme weather conditions are very complex, the high accuracy identification of them is facing severe challenges. In this article, the spectra of these three weather conditions, as well as those of clouds and the background surface, are analyzed. A monitoring model is constructed to 15 achieve the separation of fog, haze, dust storm, clouds and the underlying surface using satellite data. The monitoring results are tested based on their corresponding measurements obtained from ground stations, and indicate that the extraction of fog, haze and dust storm can reach a high accuracy .


Introduction
Fog, Haze and Dust storm are three typical extreme weather that often occur in China, especially in the areas where 20 the ecological environment is seriously damaged.(Miri et al., 2010;Zhang et al., 2013;Sallis et al., 2014;David et al., 2012;Naksen et al., 2017).Fog, haze and dust storm can cause the climate changes by affecting the radiation budget and energy balance of the Earth-atmosphere system (Quinn and Bates, 2003;Kaskaoutis et al., 2006;Elias et al., 2009;Maghrabi, 2017).When such weather happens, atmospheric visibility is sharply reduced, which seriously affects the transportation (Sisler and Malm, 1994;Deng et al., 2011).Fog，haze and dust particles can also produce 25 air pollution (Yu et al., 2011;Kang et al. 2013;Quan et al., 2014;Gao and Chen, 2016), increasing the concentration of inhalable particles and leading to respiratory and cardiovascular diseases (Thurston et al., 1994;Bai et al., 2006;Zhao et al., 2013;Gao et al., 2016;Gao et al., 2017).Detecting the fog, haze and dust storm still faces great challenge for the following reasons: (1) It is difficult to distinguish between low clouds and fog, as they have similar spectral features.(2) The remote sensing detection of haze is rarely performed, current haze detection still mainly depended on the ground observation data.(3) Previous studies did not distinguish the three extreme weather at the condition of simultaneous occurrence of them.From a wide range of monitoring, such situation often occurs, but because of their optical characteristics are very similar, 5 monitoring in this case is very difficult.
The physical characteristics of fog, haze and dust storm are much different, they have many differences in particle size, moisture content and so they show different spectral characteristics, especially in the visible and infrared wavelength.Remote sensing technology takes advantage of this optical difference to realize the identification of them.MODIS (Moderate Resolution Imaging Spectrometer) data are used to do the experiment for the detection of 10 the three extreme weather phenomena.Based on the spectral analysis of this three kinds of extreme weather and the remote sensing data characteristics, this paper construct a high accuracy model for the fog, haze and dust storm identification.

15
MODIS is a key instrument aboard the Terra and Aqua satellites.They are acquiring data in 36 spectral bands, ranging in wavelength from 0.4 µm to 14.4 µm.Two bands of 0.66µm and 0.86µm are imaged at a nominal resolution of 250 m at nadir, with five bands of visible to near infrared band at 500 m, and the remaining 29 bands at 1 km, they achieves a 2,330km swath and provides global coverage every one to two days.There are many standard MODIS data products that scientists are using to study global change.These products are being used by scientists 20 from a variety of disciplines, they are playing an important role in ecological environment monitoring and monitoring of global climate change.

2 Data preprocessing
The Level 1B data set contains calibrated and geolocated radiances generated from MODIS Level 1A sensor counts.
The radiances are in W/(m2-µm -sr).The parameters of apparent reflectance and brightness temperature are used to 25 construct the detection model.Apparent reflectance is the ratio of upwelling irradiance to the downwelling irradiance, where ref represents the apparent reflectance, E is the solar irradiance, and D is the earth-sun distance, which is generally defined as 1. cos is the cosine of the solar zenith angle.

5
The brightness temperature is a measurement of the radiance of the thermal infrared radiation traveling upward from the top of the atmosphere to the satellite, expressed in units of the temperature of an equivalent black body.
The brightness temperature is the fundamental parameter measured by thermal infrared radiation sensors, which can be obtained from Planck's formula.
where h is Planck's constant, which is defined as 6.626×10 -34 J.s; k is the Boerziman constant, which is defined as 1.3806×10 -23 J.K -1 ; c is the velocity of light, which is defined as 2.998×10 8 m.s -1 ;  is wavelength, and T is temperature.Fig. 1 is the true color image of dense fog weather, which presents as white with a smooth top structure, even texture and clear borders.Fig. 2 is the true color image of haze weather, which presents as gray with an even texture and clear borders.Fig. 3 is the true color image of dust storm weather, which presents as drab yellow feathers with a uniform top structure and fuzzy borders.In each image, cloud areas are present as white regions with variable brightness, rough top structures and messy borders.

5
Band 1~19 and 26 of the MODIS are visible and near infrared bands, the energy received by these bands is from the reflected solar radiation of objects, bands 20~25 and 27~36 are thermal infrared bands, the energy received by the satellite mainly originated from the emitted radiation of the atmosphere and the land.The difference of reflectance or brightness temperatures of fog, haze, dust storm, clouds and the underlying surface were analyzed based the sample areas selected from the MODIS images of Fig. 1, 2 and 3.The conclusions of this analysis are as 10 follows: the combination of bands 22 and 23 can be used to define fog areas, bands 2 and 8 can be used to detect the haze areas, and bands 31 and 32 can be used to identify dust storm areas.

5
-0.42~0.15,peaking at -0.33, HI of the underlying surface reflectance ranges from -0.33~0.12,fog and clouds are distributed from -0.06~0 and -0.03~0.12,respectively.The HI of haze is larger than 0.03; thus, a haze area can be identified if HI >0.03.The above combination of wavebands could be used to effectively distinguish between fog, haze and dust storm areas.However, the underlying surface effect must also be eliminated.The combinations of wavebands and bands used to separate the underlying surface are described as follows. 5 Fig. 7 shows the difference image of bands 20 and 31, in which the green dots represent fog observation stations.
Fog areas are expressed in white, cloud areas are expressed in white and black-gray, and the underlying surface is ash-black.In the brightness temperature difference image, differences in color represent differences in brightness temperatures.shows the cloud areas have the highest reflectance, followed by dust storm, the underlying surface is the lowest.Fig. 12 shows that the reflectance of dust storm ranges from 0.12 to 0.3, the underlying surface is 0.06-0.22 and the clouds is 0.12-0.66.Part of the underlying surface could be identified using a threshold of 0.12.3. Experiment and validation 5

Validation data introduction
The data used for validating the extreme weather detection method is obtained from MICAPS (Meteorological Information Comprehensive Analysis and Process System), which is a software tool for visualizing data, and is popular used by China Meteorological Administration (CMA) for weather forecast.A total of 19 kinds of data is in the MICAPS, include the ground observation data, satellite data, and so on.Table 1 is the statistical results of the accuracy for fog detection, totally, 36 ground sites were distributed in the fog area, among them, 9 sites were covered by clouds.As is shown in Table 1, the scope of the heavy fog at 9 sites covered by cloud, as well as 4 mist covered sites, were not detected.The detection accuracy of fog areas is 63.9%.
This accuracy might even higher if no cloud influence.

Instance analysis of haze detection
As is shown in Table 2, 22.4% of all sites were covered by the cloud, with an extracting accuracy of 74.1%.This accuracy could even higher if without cloud influence.
Instance analysis indicates that the multichannel threshold value method proposed in this paper could have favorable effects on the detection of fog, haze and dust storm.However, due to cloud coverage, data from under the cloud could not be detected by the MODIS sensor.Additionally, light areas were not detected, and the borders of

Conclusions
This paper uses MODIS as a main data resource in order to systematically analyze the physical and spectral characteristics of three types of extreme weather, namely, fog, haze and dust storms.The spectral differences between fog, haze, dust storm and clouds, as well as the underlying surface were analyzed, to determine the bands 5 used for detecting the typical extreme weather events.Experiments of fog, haze, dust storm detection were performed based on the MODIS data, and the evaluations were made in conjunction with the data collected from MICAPS.The conclusions can be described as follows: (1) When fog, haze or dust storm weather occur, dramatic changes occur in moisture and particle content; due to differences in their physical characteristics, the underlying surface and the three extreme weathers exhibit obvious 10 differences in the behavior of their spectra.Remote sensing detection could be performed using the appropriate bands.
(2) In some wavebands, the three extreme weathers exhibit similar cloud reflectance values and brightness temperatures.Some cloud detection bands cannot be used to distinguish between clouds and the three weathers.
Therefore, more bands are required for cloud detection in such weather condition.

15
(3) Experiments and validation shows that the detection methods for fog, haze and dust storm, developed in this research, can work well in most areas, but if they are covered by clouds, the accuracy is very poor, also this method is not good in the edge detection of the extreme weather.
Atmos.Meas.Tech.Discuss., https://doi.org/10.5194/amt-2017-306Manuscript under review for journal Atmos.Meas.Tech.Discussion started: 23 October 2017 c Author(s) 2017.CC BY 4.0 License.and can be calculated for the solar reflective bands through knowledge of the solar irradiance with the

Fig. 4
Fig. 4 is the distribution of the brightness temperatures of fog, haze, dust storm, clouds and the underlying surface in different image of bands 22, 23, it shows that the differences in cloud brightness temperatures are distributed widely between 1-11K, peaking at 5K; the differences in the brightness temperatures of the underlying surface, haze 15

Fig 4 .
Fig 4. The distribution of the brightness temperatures of fog, haze, dust storm, clouds and the underlying surface in difference

Fig 5 .Fig 6 .
Fig 5.The HI of fog, haze, dust storm, clouds and the underlying surface in a 2, 8 band combination image

Fig 12 .
Fig 11.The MODIS image of band 1

Fig. 16
Fig. 16 (a) shows the MODIS data affected by haze weather at 14:00 on February 20, 2008.Green dots is the location of ground haze observation sites, gray areas in the image is the haze distribution.It indicates that the haze weather influenced a large area of central and eastern part of China.And a total of 85 ground haze detection sites were distributed in such area, among them, 19 sites were covered by clouds, like the south of Hubei and Jiangxi 10

Fig. 16
Fig. 16 (b)  is the cloud identification and fig.16(c) is the haze detection images.These images show that the haze areas can be identified in high precision without cloud influence, however for cloud covered area, most of the haze distribution can't be detected.

Table 1 .
The accuracy of fog detection 5

Table 2 .
The accuracy of haze detection