A neural network method for monitoring snowstorm: A case study in southern China

It has been observed that low temperature, rainfall, snowfall, frost have never occurred over the past 50 years in the southern China, and weather in this area is very complex, so the monitoring equipments are few. Optical and thermal infrared remote sensing is influenced much by clouds, so the passive microwave Advanced Microwave Scanning Radiometer-Earth Observing System (AMSR-E) data are the best choice to monitor and analyze the development of disaster. In order to improve estimation accuracy, the dynamic learning neural network was used to retrieve snow depth. The difference of brightness temperatures of TB18.7V and TB36.5V, TB18.7H and TB36.5H, TB23.8V and TB89V, TB23.8H and TB89H are made as four main input nodes and the snow depth is the only one output node of neural network. The mean and the standard deviation of retrieval errors are about 4.8 cm and 6.7 cm relative to the test data of ground measurements. The application analysis indicated that the neural network can be utilized to monitor the change of snow intensity distribution through passive microwave data in the complex weather of the southern China.


Introduction
Snow cover significantly influences the Earth′s surface energy balance and acts as the frozen storage in water balance.Snow parameters (especially snow water equivalent) are important variables for flood forecasting and the management of water resources, such as reservoir management and agricultural production activities (Jiang et al., 2007).Many experiments like Cold Land Processes, World Climate Research Programme, Cli-mate and Cryosphere, Global Energy and Water Cycle Experiment have been made by many scientific research organizations, but people still do not have sufficient capacity to predict and assess the influence of snowstorm.Snow cover and snow water equivalent (SWE) data are usually monitored through traditional meteorological stations, which are poorly distributed globally in the point form (Robinson et al., 1993).However, the number of meteorological stations is often not enough and the distribution is not at random, especially for moun-tainous region.
The remote sensing technology provides a chance for us to quickly and efficiently obtain a large area of snow cover, snow depth, and many other parameters.It has been about 30 years to use remote sensing technology for monitoring snow.The methods to estimate the snow cover for visible and near infrared sensors, such as Moderate Resolution Imaging Spectroradiometer (MODIS) and National Oceanic and Atmospheric Administration/ Advanced Very High Resolution Radiometer (NOAA/ AVHRR), are increasingly mature, but the estimations of snow depth and snow water equivalent are still at the development stage (Chang et al., 1987;Tait, 1998;Kelly et al., 2003;Langlois et al., 2011;Durand and Liu, 2012).The brightness temperature of frequencies at 3-90 GHz is sensitive to snow crystal characteristics, snow density, and snow water equivalent (Pulliainen et al., 1999;Wiesmann and Mätzler, 1999;Macelloni et al., 2001;Tsang and Kong, 2001).The influence for each frequency is different under different conditions.For example, the emission from dry snow is mainly affected by underlying soil dielectric and roughness properties at lower frequencies.The emission is sensitive to snow water equivalent and snow particle size at higher frequencies because the volume scattering by snow particles becomes important (Mätzler, 1996).So Microwave remote sensing has been proved to be very effective in obtaining snow parameters, especially for snow depth and snow water equivalent (Foster et al., 1984;Shi and Dozier, 2000;Derksen et al., 2005;Foster et al., 2005;Macelloni et al., 2005;Pulliainen et al., 2006).In particular, passive microwave remote sensing has become an important data source of snow cover information because of all-weather imaging capabilities, rapid revisit time, and the ability to derive quantitative estimates of SWE (Derksen et al., 2000;Jiang et al., 2007;Langlois et al., 2011;Durand and Liu, 2012).Many researches have been done by using passive microwave remote sensing to obtain snow parameters and have obtained large progress in often snowy areas (Foster et al., 1997;Pulliainen and Hallikainen, 2001;Kelly and Chang, 2003;Roy et al., 2004;Tedesco et al., 2004).
Although low temperature and snow disasters are common meteorological disasters during the Chinese winter, especially in the Qinghai-Tibet Plateau, northern Xinjiang, Inner Mongolia and Northeast China, they rarely occur in the southern China (Dong et al., 2001;Qin and Sun, 2006;Zhang et al., 2009).But in 2008, there were four consecutive high-intensity snowfalls for more than 20 days which significantly struck the transportation, communication and electricity transmission lines.It has never occurred over the past 50 years in the southern China and the weather is very complex in this area, but little research is done on low-temperatures and snow disasters in the southern China.Optical and thermal infrared remote sensing data (like MODIS) is influenced much by clouds, and the resolution of passive microwave is low but high coverage which is suitable for monitoring disasters with large areas, so the passive microwave Advanced Microwave Scanning Radiometer-Earth Observing System (AMSR-E) data are the best choice to monitor and analyze the development of the disaster.In order to improve the monitoring accuracy for snowstorm under complex weather conditions, this paper makes a case study by using microwave remote sensing data in South China. .

Study area
The climate differs from region to region because of the extensive and complex topography in China (Dong et al. 2001).The Changjiang (Yangtze) River serves as China′s official dividing line between north and south, to some extent, it is also the dividing line of climate because the water vapor content from north and south usually meets here.The water vapor content in the southern China is higher than that in the northern China because of the influence of the Pacific and Indian Oceans.Clouds are very thick in the southern China, especially in Changjiang River Valley, which makes radiation of microwave in the region more complex and different from the other regions of China.The area of severest cloud coverage suffered from low-temperature, frost, heavy rainfall and snowfall in early 2008 are selected as the research region, which mainly involves Hunan, Hubei, Anhui, Guizhou, Guangxi, Guangdong, Jiangsu, Jiangxi, Shanghai, Chongqing, Zhejiang and Fujian (Mao et al., 2009).The range of study area is marked blue in Fig. 1.

Data and processing
AMSR-E is a passive microwave radiometer, which observes atmospheric, land, oceanic, and cryospheric pa- rameters, including precipitation, sea surface temperature, ice concentration, snow water equivalent, surface wetness, wind speed, atmospheric cloud water, and water vapor content (Mao et al., 2007a;2007b).The AMSR-E Level 2A product (AE_L2A) contains brightness temperatures (TBs) at 6.9, 10.7, 18.7, 23.8, 36.5, and 89.0 GHz, which are resampled to be spatially consistent, and it is available at a variety of resolutions that correspond to the footprint sizes of the observations (56, 38, 24, 21, 12, and 5.4 km, respectively).The data can be downloaded freely from National Snow and Ice Data Center (http://nsidc.org/data/ae_l2a.html).In this study, the AMSR-E data with the resolution of 24 km × 24 km were used to monitor the change of snow area and the distribution of snow depth with the time.The brightness temperature can be obtained by Equation (1).
where BT i is brightness temperature of frequency i; Offset and Scale_factor can be obtained from the head file; DN is the value of the remote sensing image.
The study area has 818 meteorological stations covered in the southern China and parts of them are shown in Fig. 1.The ground measurement data in meteorological stations are collected to make comparisons and evaluations.

Neural Network
Neural networks learn from training data which need not be programmed to follow a specific mathematic equation.Many people (Tsang et al., 1992;Tzeng et al., 1994;Faure et al., 2001;Tedesco et al., 2004;Mao et al., 2007b;2008a;2008b;2010) used the neural networks to perform inversion for geophysical parameters and obtained good retrieval results.The dynamic learning neural network (DL) (Tzeng et al., 1994) is selected to solve retrieval problems, which uses the Kalman filtering algorithm (Bletsas, 2005) to increase the convergence rate in the learning stage and enhance the separate ability for highly nonlinear boundaries problem.The initial neural network weights are set to be small random numbers (-1, 1).The Kalman filtering process is a recursive mean square estimation procedure.Each updated estimate of neural network weight is computed from the previous estimate and the new input data.The weights connected to each output node can be updated independently.The DL quickly achieves the required root mean square (rms) error in just a couple of iterations, and the result trained by NN is very stable.Thus the root error threshold is often set to be 0.001 and the epochs of iteration is two.More advantage of dynamic learning neural network than BP/FL (fasting learning) neural network can be found in reference (Tzeng et al., 1994).

Results and Analyses
The AMSR-E data from January 10, 2008 to February 10, 2008 are downloaded from National Snow and Ice Data Center of US.One hundred and fourteen meteorological stations in the southern China are selected as ground data collection sites.A program has been made to read AMSR-E brightness temperatures from January 10, 2008 to February 10, 2008 by using the longitude and latitude as control condition.We collected 1524 data sets, and divided them randomly into two groups.The training data are 1101 sets and the testing data are 423 sets.The difference of brightness temperatures of TB 18.7V and TB 36.5V , TB 18.7H and TB 36.5H , TB 23.8V and TB 89V , TB 23.8H and TB 89H are made as four main input nodes and the snow depth is the only output node of neural network.After trial and error, we set the number of hidden-nodes from small to large, and selected the smallest of retrieval error for testing data (Table 1 is just part of data).The mean and the standard deviation of retrieval error are about 4.8 cm and 6.7 cm relative to the ground measurement data when the number of the hidden layers is 2 and the number of hidden nodes is 350 in each layer.The neural network trained above was used to estimate the change of snow-depth from January 10th to February 10th 2008.The part of retrieval results is shown in Fig. 2, and the different color presents how much depth of snow.The disaster area reached maximum between January 28th and February 1st, 2008, which are consistent with the ground measurement.Figure 2c shows that the disaster is a very serious in Hunan, Guizhou, eastern of Sichuan, most part of Guangdong, Guangxi and Jiangxi provinces.The snow melted very quickly because weather changes were rapid in the southern China.A few values of estimation are anomalies with a big estimation error because the weather is too complex in the southern China, so the scale of snow depth in Fig. 2 is adjusted according to the data from meteorological stations.Table 2 is part of disaster area information of the southern China obtained from meteorological stations.. Shown from Fig. 2 and Table 2, the disaster area estimated from AMSR-E data is consistent with the data obtained from ground meteorological stations.
Although the weather is very complex in the southern China, the snow is so large that the snow depth is relatively uniform in some places.Figure 3 is a comparison between estimation from AMSR-E data and ground measurement data from January 31, 2008 to February 10, 2008.The square of correlation coefficient is about 0.7, and mean relative error is about 5.2 cm. Figure 2 indicates that the estimation error is large at the verge of sea and lake because the radiation mechanism is different between water surface and land surface.There are two main reasons (Mao et al., 2007a).The first is that there exists certain water like large lake in some AMSR-E pixels because the radioactive mechanism of water is different from others (soil and vegetation).The second is that there exists certain major rainfall in some AMSR-E pixels.On the other hand, the resolution of passive microwave data is very low, so there are many mixed pixels.In order to improve estimation accuracy,

Discussion
It is very difficult for visible and thermal remote sensing to obtain ground information because there are much cloud and rainfall in the southern China, but microwave remote sensing can overcome this difficulty.The microwave radiometer measures the thermal emissions of the ground which transfer from the ground through the atmosphere to the remote sensor.According to the Raleigh-Jeans approximation for Planck function, the intensity of radiation observed by a radiometer can be simply written as Equation ( 2) (Mao et al., 2008c).(1 )(1 ) ( 1) where p stands for horizontal (H) or vertical (V) polarization; μ = cosθ; ε p is emissivity; τ (equivalent canopy optical depth) and ω (the single scattering albedo) are two important parameters that characterize the absorbing and scattering properties of vegetation; T s is land surface temperature; T c is average temperature of vegetation; T bp (τ, μ) is brightness temperature of radiation emitted by canopy at an angle θ; t is transmittance of atmosphere; a T  is upwelling average atmosphere temperature; a T  is downwelling average atmosphere temperature.The centimeter wave band is influenced little by the atmosphere, and the transmittance (t) of microwave is very high and approximates to 1 even if the water vapor content is about 5 g/cm 2 in atmosphere (Ulaby et al., 1986), so the influence of atmosphere is very little and Equation ( 2) can be simplified to Equation (3).
T c is usually assumed to be equal to T s in soil covered by vegetation (Paloscia and Pampaloni, 1988; Njoku et al.,   2003).For ground surface covered by snow, the emission of surface in the southern China is more complex than the description above.To determine the total emission from a snow layer above ground, the geometry of ground should be considered, which is shown in Fig. 4. The total emission source (T s ) within a layer can be separated into three components: the total upwelling emission (T us ), the total downwelling emission (T rds ), and the ground emission into the layer (T gs ).The ground emission is determined by the temperature and emissivity, which depends on the soil moisture and roughness of snow covered land.The emissivity is mainly influenced by dielectric constant which is a function of physical temperature, salinity, water content, soil texture and other factors (Dobson et al., 1985;Hallikainen and Jolma, 1986;Jiang et al., 2007).However, as shown by equations ( 2) to ( 5) and Fig. 4, the emission of ground surface is very So it is very difficult to accurately estimate the snow depth from passive microwave remote sensing because a single frequency thermal measurement has at least two unknowns (emissivity and temperature) which is a typical ill-posed inversion problem.On the other hand, the dry snow emits considerably less microwave radiation than soil, and the brightness temperature of snow is inversely related to the snow water equivalent.When snow starts to melt, emission will significantly increase because water droplets absorb and re-emit rather than scatter microwave radiation (Foster et al., 2005;Jiang et al., 2007).So the Fig. 4 Emission of ground surface covered snow retrieval is very complicated because different combination of snow, soil moisture, roughness, and land surface temperature may produce the same brightness temperature for single frequency.The relationship among emissivity, snow (sw), soil moisture (sm), roughness and temperature (T s ) can be depicted as Equation ( 6).The weather is very complex in the southern China.There are different types of cloud and rain in different regions.The rainfall, snowfall, snow and frost exist at the same time, and the status of snow changes with temperature.The different status of snow at different frequencies influences radiation of microwave differently.The emission of dry snow is considerably less than soil, so the brightness temperature of snow is inversely related to the snow water equivalent.When snow starts to melt, emission will significantly increase because water droplets absorb and re-emit rather than scatter microwave radiation (Foster et al., 2005;Jiang et al., 2007).The emission at lower frequencies from dry snow is mainly determined by the underlying soil dielectric and roughness properties.The emission at higher frequencies is more sensitive to snow water equivalent and snow particle size since the volume scattered by snow particles becomes important (Mätzler, 1996;Jiang et al., 2007).The general retrieval method is not very suitable for complex weather conditions (lowtemperature, freezing, heavy rainfall and snowfall) of the southern China in early 2008.Another important reason is that most pixels are mixed for coarse spatial resolution of passive microwave (24 km × 24 km).The retrieval equations are so complex that a solution is often difficult to derive.Fortunately, the neural network is much different from conventional algorithm requiring that the inversion algorithm be known exactly.Neural network owns function approximation, optimization computation and classification ability which is composed of a large number of highly interconnected processing elements (neurons) working in parallel to solve a specific problem.All analysis indicates that neural network is more competent for retrieving snow depth in complex weather.

Conclusions
With the increasingly extreme climate changes, the monitoring of snowstorms in the southern China should be strengthened because the region has low levels of temperature, rainfall, snow, and frost at the same time.It is of urgent matter to monitor disasters via remote sensing images.Generally, for the geophysical parameters retrieval from passive microwave remote sensing data, it is very difficult to build a set of general retrieval equations due to many nonlinear and poorly understood factors.The analysis indicates that traditional methods can not depict all conditions well and is not suitable for the southern China.
In contrast to conventional methods, the neural network does not require that the relationship between input parameters and output parameters be known, since it is directly determined by the training data.Due to the fact that NN simultaneously owns function approximation, classification, optimization computation, and selfstudy ability, we can utilize neural network by using reliably measured ground data to estimate the snow depth.The mean and the standard deviation of retrieval errors are about 4.8 cm and 6.7 cm comparing with the test data of ground measurement data.The application analysis indicates that the neural network can be utilized to monitor changes in snow cover and the intensity distribution from passive microwave data.In order to improve estimation accuracy, more research should be made further for mixed pixels covered by water and other influence factors, such as topography, vegetation, mixed pixel.

Fig. 1
Fig. 1 Location of study area and distribution of part meteorological stations

Table 1
Summery of retrieval error

Table 2
Disaster area in the southern China in 2008 research should be made further for mixed pixels covered by water and other influence factors, like topography, vegetation, mixed pixel.
Fig. 3 Comparison of snow depth between estimation from AMSR-E data and ground measurement data more