Relationship between the ground surface temperature by NOAA-AVHRR and environmental factors

Relationship between the ground surface temperature by NOAA-AVHRR and environmental factors

Ryuzo Yokoyama, Chang Ming Zhou, Sumio Tanba
Department of Computer Science
Faculty of Engineering, Iwate University
4-3-5 Morioka, Iwate Japan 020

Abstract
The relationship between the ground surface temperature observed by AVHRR and the environmental factors of elevation cosg, vegetation index, latitude and albedo were investigated by the regression analysis. In the test site of about 600kmx200km in the northern Honshu of Japan, sample points were specified at lattices with 5km distance. In an April image, no single variables had high correlation coefficients, but in June image, the elevation, the vegetation index and the albedo showed correlation coefficients higher tha 0.6. The final correlation coefficients in the multi-variate regression were 0.540.85. Large residues appeared in seaside regions and basins.

Introduction
The ground surface temperature is one of the most fundamental variables of the natural environment. As shown in fig,1 it is determined by the balance of heat fluxes at the surface, which are participated by the five components of

R_e : Long wave radiation from the ground
R_L : long wave radiation to the ground
R_s : short wave radiation to the ground
LE : flux of latent heat,
|| : flux of sensible heat,
G : energy flux into the ground

The advent of the remote sensing has provided a new method of temperature observation. Being a spatial information at an instant, remotely sensed ground temperature images have been applied to the investigations of the environment, e.g. heat island, thermal inertia, land cover classification etc.

In this paper, we investigated the relation between the ground surface temperature observed by the NOAA-AVHRR and other environmental factors by the regression analysis. The environmental factors, which include the elevation, the vegetation index, cos, the albedo and the latitude, are selected as the variables such that they are closely related to the components of the heat balance variables such that they are closely related to the components of the heat balance variables such that they are closely related to the components of the heat balance and can be systematically read out from existing data sets by computer. Cos means the angel between the normal vectors of the ground surface and the direction of the sun light incidence at the data acquisition.

Figure 1: Heat flux balance at the ground surface.

Preparation of data set
As shown in Fig. 2 the test site was specified to the Tohoku district of Japan with an area of 600x200km. Its east side and west side face to the pacific ocean and the sea of Japan, respectively. Oou mountains, of which highest part is about 2,000m, runs from north to south in the central part. The site has a rather complicated geography because there are many rivers from the mountain areas formulating valleys, basins and plains.

The analysis was applied to the two AVHRR image data in Table 1. Fig. 2 and Fig. 3 show the pseudo-color and the temperature displays of the images. As the test site has been covered by traveling high pressure before the data acquisitions, those images represent typical data of the spring and the early summer under the stable clear sky conditions.

Table 1. Acquisition time of AVHRR data used in the analysis

	Local time	Date	#NOAA
Image-A	15:12	1988.4.17	NOAA-9
Image-B	14:03	1986.6.12	NOAA-9

Figure 2: The map of Japan. The test site is the region inside the rectangle.

Figure 3: Pseudo color displays of AVHRR data used in the analysis.

This district was divided into square meshes with 5km distance, and sample points were specified at the center of each mesh. At each sample point, values of each variable were read out or calculated as follows .

ST (AVHRR ground surface temperature) : After the geometric correction of the original images, the temperatures were calculated by the method in Lauristein (1979). Then they were calibrated by the SST estimation function of Yokohama (1988).

ELE (elevation): This was read out from the digital geographical data set of Japan (DGDSJ). As the mesh size of DGDSJ I s250m and the ground resolution of AVHRR is 1.1km ELE was calculated as the mean of 5x5 neighborhood pixels in DGDSJ

Cosg : This was calculated from DGDSJ and the sun position.

VI (Vegetation index ) : This was calculated from Ch. 1 and Ch. 2 data of AVHRR as

VI = ch.2 - ch.1 / ch.2 + ch.1 x 100
Lat (latitude coordinate) : This was read out from DGDSJ.

AL (albedo in the visible ch.1 of AVHRR) : This was calculated by following to the definition in Lauristein (1979).

In the preparation of the final data set, sample points accepting one of the following conditions were exempted

A sample point of which neighboring points with 1.0 km were covered by cloud, snow or noises.
A sample point of which neighboring points within 1.0 km include a point in water.
A sample point at which values in the neighbor hood points are very variable. The threshold value were determined from the histogram of the standard deviation of neighborhood points.

Table 2 shows the statistics of the final data set. Since the high elevation areas in image A had been covered by snow, its total number of final sample points and the mean of ELE is smaller than those of Image-B because most of deciduous trees, grasses and crops had not grown their elves yet in the middle of April.

Table 2: Statistics of the final data set for the regression analysis

Image data	#of samples	ELE [m]	Cosg	VI	LAT	AL	ST [°C]
Image-A	1345	195.14 258.89	0.06 0.54	5.17 21.77	1.11 38.98	0.79 6.25	2.10 19.53
Image-B	1712	266.50 377.17	0.05 0.78	11.81 51.83	1.14 38.88	0.92 6.56	3.02 26.81

Results of regression analysis
The regression analysis was applied by assuming the object variable to be ST. The correlation coefficients and the standard errors (standard deviation of residues0 were shown in Table 3.

Table 3: Correlation coefficients and standard errs in the regression analysis

Explanation variables	Result for Image-A	Result for Image-B	Explanation variables	Result for Image-A	Result for Image-B
E, G, V, L, A, E,G, E,V, E, L, E, A, G, V, G, L, G, A, V, L, V, A, L, A, V, A, L, A,	-0.362/1.956 0.287/2.010 -0.394/1.928 -0.126/2.081 0.266/2.022 0.439/1.886 0.463/1.860 0.403/1.921 0.406/1.918 0.468/1.855 0.307/1.997 0.370/1.060 0.402/1.921 0.403/1.921 0.289.2.009 0.403/1.921 0.289/2.089	-0.599/2.422 0.392/2.783 -0.768/1.935 -0.002/3.024 0.674/2.235 0.645/2.311 0.789/1.857 0.619/2.374 0.773/1.920 0.784/1.879 0.392/2.783 0.710/2.129 0.771/2.926 0.785/1.875 0.704/2.149 0.785/1.875 0.704/2.149	E,G,V, E,G,L, E,G,A, E,V,L, E,V,A, E,L,A, G,V,L, G,V,A, G,L,A, V,L,A, E,G,V,L, E,G,V,A, E,G,L,A, G,V,L,A, E,G,V,L,A,	0.518/1.795 0.467/1.857 0.468/1.856 0.481/1.840 0.469/1.854 0.437/1.888 0.473/1.850 0.479/1.843 0.383/1.939 0.410/1.915 0.531/1.780 0.527/1.786 0.490/1.831 0.483/1.839 0.537/1.772	0.801/1.812 0.991/2.271 0.788/1.862 0.798/1.825 0.810/1.773 0.819/1.736 0.786/1.871 0.799/1.821 0.735/2.053 0.795/1.836 0.808/1.782 0.820/1.731 0.829/1.694 0.808/1.785 0.842/1.631

In the single variate regression, the correlation coefficients of ELE, VI and LAT are negative. Those might be reasonable because it si well known that higher elevation and latitude provide cooler climate, and more active transpiration is expected in the regions with higher VI. On the other hand, the correlation coefficient of Cosg is positive since under the clear sky, more sunshine arrives to the ground with larger cos. It may sound strange that the correlation coefficient of AL is positive. But in the two images, the correlation coefficients between VI and AL were less than -0.75. That is, the area of urban districts, bare soils, crop fields and grasslands had higher AL values and the forest area had lower AL values. The higher correlation of AL to ST can be similarly understood as the case of VI. Cukovishi 91987) observed the negative correlation of AL to ST in Senegal.

In the single regression analysis of Inage-B, ELE, Vi and AL showed high correlation to ST. Those should be dominant variables in the ground surface temperature formulation as stated in the above. Correlation coefficients of those in image-A, however, were kept low. This might come from the small dynamic range of ST. the growth of vegetation in Image-B provided a larger dynamic range of ST. The Cos and LAT showed small correlation coefficients in both images.

In the multi-variate regression, image-B showed higher correlation coefficients than Image-A. This is due to the wider dynamic range of ST. By increasing the number of explanation variables, the correlation coefficients were improved. Figure 5 shows the scatter diagrams between the AVHRR temperature and the estimated temperature with higher correlation coefficients.

Figure 4: Temperature images used in the analysis.
The brighter a region is, the higher is its temperature.

Residue images
It will be interesting to observe the differences between the AVHRR temperature image and the estimated image. The results are shown in figure 6 in which the regression functions in fig. 5 were applied. Except the regions of exempted sample points (mostly snow covered areas), characteristics residues appeared in specific regions. Minus large residues, which mean the AVHRR temperatures are lower than their estimated ones, were concentrated is seaside regions of which ground surface temperature could be affected by the sea temperature. The tendency is more prominent in Image-B when the air temperature had been warmed up in the middle of June but the sea temperature was still cold.

Plus large residues mostly appeared to basins. The AVHRR temperatures might have become higher than the estimated under the clear sky.

Figure 5:

Figure 6: Temperature difference images between
the AVHRR observed and the estimated.

Conclusion
For Image-B of June, the AVHRR ground temperatures were effectively estimated with the correlation of coefficient of 0.84. The dominant variable in the estimation were the elevation and the vegetation index. The albedo had a high correlation but it is dependent upon VI. For image-A of April, however, the final correlation coefficient was in a level of 0.54 and prominent variable were not existed. Large plus residues mostly appeared in seaside regions, and large minus residues appeared in basins.

More studies are expected to the images in other season to investigate the ground surface temperature formulation under various environment conditions.

References

Lauristein, R.L., Nelson, G.L. & Porto, F.W. "data extraction and calibration of TIROS-A / NOAA radiometer", NOAA Tech. Memo NESS 107, 1979.
Yokoyama, R. & Tanba, S., "Estimation of Sea surface temperature via NOAA-AVHRR sensor : comparison with sea truth data by fixed buoy", Proceedings of IGARRS"88 symposium., pp. 275-280, 1988.
Vukovbich, F.M., Toll, D.L. & Murphy, R.E, "Surface temperature and aldebo relationships in Senegal derived from NOAA-7 satellite data" remote sensing of environment 22, pp413-421, 1987.