^{1}

^{1}

^{*}

The objective of this paper is to utilize images of spatial and temporal fluctuations of temperature over the Earth to study the global climate variation. We illustrated that monthly temperature observations from weather stations could be decomposed as components with different time scales based on their spectral distribution. Kolmogorov-Zurbenko (KZ) filters were applied to smooth and interpolate gridded temperature data to construct global maps for long-term (≥ 6 years) trends and El Nino-like (2 to 5 years) movements over the time period of 1893 to 2008. Annual temperature seasonality, latitude and altitude effects have been carefully accounted for to capture meaningful spatiotemporal patterns of climate variability. The result revealed striking facts about global temperature anomalies for specific regions. Correlation analysis and the movie of thermal maps for El Nino-like component clearly supported the existence of such climate fluctuations in time and space.

The general feature of the global climate variability is an important topic for climate researchers. The past decades had seen tremendous progress in developing consistent database of meteorological observations covered long time periods [1-4]. These works generally involved assimilating, gridding and interpolating surface climate records to form a unified temperature field covering the land area of the Earth. The availability of these datasets facilitates revealing and visualizing the climate movement at the global level. By utilizing one of the datasets, we built a fine-resolution description of the mean states and space-time fluctuations of the global climate to illustrate the basic pattern of climate movement. Compared to other existing “reanalysis” works [

Our work focused on surface temperature, the most important role in investigating the global or regional climate change [

From the statistical perspective, to build the profile of global climate is a typical space-time modeling task. The major challenge is the complex spatiotemporal dependencies over multiple time and spatial scales [9,10]. The presence of various scales of motion in time and space complicates the analysis and interpretation of the data [11,12]. Therefore, the first step to solve the problem should be decomposing meteorological data into components that we are interested in [

The long-term trend and El Niño-like component usually have relative smaller amplitudes compared with seasonal temperature changes and cannot be observed directly. To separate out those data components, we introduced the Kolmogorov-Zurbenko filter (KZ) [

The following sections describe the construction of the long-term component and El Niño-like movement of surface climate over global land areas for the period of 1893 to 2008. We will demonstrate how the long-term component movie facilitates capturing the spatial pattern of energy input and distribution over the globe land, and how the movie for El Niño-like movement helps to understand the short term temperature fluctuation and its spatial feature. The correlation analysis over these climate components revealed striking spatial/temporal correlation patterns and will also be discussed.

The data source for our study is the station-based monthly mean temperature data (version 3) from the Global Historical Climate Network Dataset (GHCN) (http://www. ncdc.noaa.gov/ghcnm/v3.php). GHCN updates this dataset daily to accommodate the newest observations; and all the data have been adjusted for homogeneity and quality control [

GHCN collects monthly mean temperature records from about 7280 stations around the world. However, these weather stations are not evenly distributed. More specifically, there are more stations in developed countries or areas with high population density. It may enlarge the bias caused by civilization [

The gridded monthly temperature data can be decomposed temporally and spatially. The following sections will address the separation of major spatial variations in the first place. After that, we will describe the method to identify and generate temporal components based on spectral analysis.

Raw temperature data has strong variations along latitude and altitude. To reasonably visualize long-term global fluctuation in space, we need to remove these effects. Following paper [

where is the sea-level long-term mean temperature on latitude y. The cosine-squared term in this formula can be explained by the product of two cosine factors. The first one is the energy density for a given sunlight beam spread on unit ground square—suppose the sun is strait up on the equator (the annual average position of the sun)—this value is proportional to the cosine of latitude. The second cosine factor is the reverse of distance in the atmosphere passed through by the sunlight before it reaches the ground. The distance is proportional to the reverse of cosine of latitude, and the double reverse determines the energy volume for a sunlight beam carried to the ground and acts as the second factor of cosine.

Temperature pattern along altitude is another spatial variation need to be addressed. Theoretically, we can convert station temperature to sea-level potential temperatures if we know station altitude and lapse rate (i.e., the rate at which temperature declines with altitude near the ground surface). However, lapse rate varies seasonally, diurnally, and regionally due to its dependence on humidity, pressure, topography, albedo, etc. The notion of constant lapse rate is an approximate description for average situation over a relative large time-space scale.

As a common practice in this area, the constant lapse rates can be calculated by regressing station temperature data on station latitudes and altitudes [7,18,19]. However, this method tends to underestimate the lapse rate value [

Beside the cosine-square pattern and altitude effect of global temperature distribution, there are several other factors also need to be controlled for. First, temperature in Antarctic is lower than the value predicted by the cosine-square law, and the plateau area in Antarctic has much higher lapse rate compared with other areas. The lapse rate for tropical region near equator also needs special treatment, although it is not as high as in Antarctic [8, 21]. Additionally, temperature in the south hemisphere tends to be slightly lower than the same latitude area in the north hemisphere. To address all of these spatial variations and interactions, the linear regression model in Equation (1) was extended as Equation (2).

where is the sea-level mean temperature on latitude y and altitude l. S and E is the dummy variable for Antarctic (y < 70˚) and equator (−10˚ < y < 10˚), respectively. Symbol “:” represents interaction between variables. Here, is used to adjust the average temperature in south hemisphere; –a_{2} is the average lapse rate, –a_{5}, –a_{6} and –a_{7} are the lapse rate adjustments in south hemisphere, Antarctic and equator regions. The spatial component represented by estimated coefficients (a_{0} to a_{7}) and related variables will be used to adjust the station observations when we apply KZ filters to generate long term singles. The total R^{2} for Equation (2) is about 0.95. The cosine-square term contributed most of the R^{2} (0.82); altitude added another 0.1 on this base; all other terms and interactions provided some extra improvement. This fact clearly shows that the cosine-square law is the most dominating factor for temperature distribution on the Earth surface. Still cosine-square law alone makes strong exaggerations in the images of elevated areas, so clear understanding of fluctuations of global temperature in space requests to include altitude factor.

This step allows us to get rid of non-interested spatial variations and uncover the long term temperature fluctuation patterns in space and time. In the result section, we will illustrate thse patterns as thermal maps of long term temperature anomaly for different area and time period.

On the time dimension, the manifest feature of monthly temperature data is the seasonal fluctuation with a cycle of 12 months. Usually the annual movements are more than 50 times stronger than signals in other frequency. Similar to the spatial variations, we need to remove this dominant variation to uncover other temporal components. The spectral feature of those temporal components therefore needs to be identified first. This can be done systematically based on spectral analysis with bootstrapping on KZ periodogram (KZP) [

KZ periodogram has strong power to separate frequencies of signals and smooth out noises. It also has the advantage to eliminate the impact of nonstationarity [

1) Randomly select 3000 stations over the globe;

2) For all stations with 50+ year records, calculate 5% DZ smoothed KZ periodogram over (0, 0.06);

3) Evenly divide (0, 0.06) as 133 bins, calculate the mean values of smoothed KZ periodogram on the 133 bins as the global raw periodogram;

4) Smooth the global raw periodogram with 7% DZ smoothing level.

The final periodogram (

We utilized Kolmogorov-Zurbenko filter (KZ) [15,22-24] as the interpolation model [_{1}, m_{2}, ···, m_{w}), k iterations of the KZ operation is represented as:

On the time dimension, m and k should be selected according to the spectra distribution of the data. Based on the average periodogram of station data (

If we set the cut-off level as half of the amplitude of output signal, the cut-off frequency for KZ_{25,5} is 0.0114 c/m [

To generate the El Niño-like component, consider the combination of KZ filters as following:

The cut-off frequencies for this equation are 0.01736 (c/m) on the left side, and 0.05556 (c/m) on the right side, corresponding to 4.8 years and 1.5 years, respectively. The leaking on annual frequency is only 0.26% and is negligible. Equation (5) will work well for El Niño-like signal on each grid cell.

Please note that we didn’t use in Equation (5). This means that the cut-off frequencies for (4) and (5) are different, and the frequencies within this range (4.8 to 7.3 years cycle length) wouldn’t be included in neither E nor G. The purpose of this design is to prevent the mixing of long-term and El Niño-like movement. Since this frequency range is not the common spectral component of global temperature records (

Next we will generate the final global long-term component G and El Niño-like component E by spatially interpolating the results of the previous section with KZ filters. In Section 2.2 to 2.4, we had already got rid of the major spatial variation by removing the long-term temperature latitude pattern and altitude effect. Now we can treat the data as isotropic and spatially smooth it with same parameters. This means that the spatial filter evenly smoothed temperature records of all the grids in space, including grids with missing data. Therefore, for the result components, the investigated temperature is proportional to physical energies in time and space.

Suppose G and E are spatially smoothed signals of G and E, we have:

where E is as in Equation (5), but usually will be enlarged to counterbalance its amplitude attenuation. Since KZ is linear filter, it is no problem for operations in (6) and (7). Here, the spatial smoothing parameter is 3˚ × 3˚ iterated 5 times, corresponding to a critical region of [15,26]. Considering the average correlation coefficient of temperature anomaly remains above 0.5 to distances of 1200 km for most latitudes [

For both long-term component G and El Niño component E, we had made movies for their thermal plots evaluated over time. This was implemented with slides generated by R lattice package on global map, aimed to visualize their correlation pattern over time and space, and facilitate capturing important events of global climate change.

We applied correlation analysis on the generated long term component G and El Niño-like component E with a re-sampling scheme:

1) For each pair of randomly selected grid points, the correlation coefficient for their time series data over common time period was calculated;

2) Draw scatter-plot for the distance-correlation relationship based on more than 10,000 samples; same for angle-correlation relationship;

3) Applied KZ filter on distance-correlation data with parameter m = 300 and k = 3; plotted the relationship;

Since we had removed the major variation on spatial dimensions for G and E, the correlation patterns are expected to be the same on all directions. Correlation analysis results verified this assumption as well as the spatial smoothing parameters used in the previous section.

We checked the spectral structure of component G and E as a verification of our design. KZ filters separated the two components as desired, and there is no signal leaking around annual frequency.

term component is spatially smoothed; its mean value is slightly different from the mean of raw data.

The spatial autocorrelation plot in