The root-mean-square error (RMSE) associated with NCI and cloud amount ranges from 0.21 to 0.30 with an averaged value of 0.25. Figure 3a shows the R-squared value between daily NCI values and the observed cloud amount in each pixel of the whole study area. Figure 4a shows the cumulative distribution function of R-squared values in the study area. At more than 80% pixels, the R-squared value is larger than 0.36 (i.e. correlation coefficient equals 0.6). The pixels where the R-squared value is larger than 0.49 (i.e. correlation coefficient equals 0.7) occupy 13% of the study area. The RMSE associated with NCI and SCI ranges from 0.16 to 0.26 with an averaged value of 0.20. Figure 3b shows the correlation coefficient between daily NCI values and the observed SCI in each pixel. Figure 4b shows the distribution of R-squared values in the study area. At more than 90% pixels, the R-squared value is larger than 0.36 (i.e. correlation coefficient equals 0.6). The pixels where the R-squared value is larger than 0.49 (i.e. correlation coefficient equals 0.7) occupy 59% of the study area.
| [width=]Cld_sun_H.eps |
| [width=0.8]Distribution.eps |
Since NDVI is sensitive to land cover, we might expect the land cover will affect NCI much. Comparing Figure 3 with Figure 1, our study indicates that the relationships between NCI and observed cloud indices retain close over most of the well-covered land uses. The R-squared values associated with satellitic remote sensing data and ground observation derivated cloud indices are shown in Table 2. The R-squared values associated with NCI and cloud amount range from 0.26 to 0.49 with a grand averaged value of 0.42 at the 7769 pixels in the study area. The USGS land use wooded tundra, mixed shrubland/grassland and water bodies are subject to the lowest R-squared values, with averaged values of 0.26, 0.27 and 0.35 respectively. The standard deviations of R-squared values over water bodies and wooded tundra are substantially larger than the standard deviations over other land uses. This indicates that the relationship is more uncertain over water bodies and wooded tundra. The land uses which are relatively well-vegetated, such as deciduous broadleaf forest, cropland/woodland mosaic, cropland/grassland mosaic, savanna, mixed forest, irrigated cropland and pasture, dryland cropland and pasture, deciduous needleleaf forest, and shrubland, are subject to R-squared values higher than 0.4. The R-squared values associated with NCI and SCI range from 0.33 to 0.57 with a grand averaged value of 0.49 in the study area. The lowest R-squared value again occurs over the USGS land use wooded tundra. The standard deviations of R-squared values over water bodies and wooded tundra are the largest two, indicating the uncertainty over these land uses. The R-squared values remain higher than 0.5 over the well-vegetated land uses which are mentioned above. The lower R-squared value in wooded tundra land suggests that frozen surface soil and snow cover might disturb the relationships between NCI and observed cloud indices.
Figure 3 shows that the poorest relationships for both cloud amount and SCI occur in the Qinghai-Tibet Plateau where the land use wood tundra is. The R-squared values against altitude at the 7769 pixels in the study area are shown in Figure 5. In Figure 5a, the R-squared values between NCI and SCI are about 0.45 when the altitude is less than 500 m. After a slight decrease to 0.40 when the altitude is between 500 m and 3,500 m, the R-squared values decrease sharply to 0.20 when the altitude reaches 4,500 m. In Figure 5b, the R-squared values between NCI and cloud amount vary mainly in the range of 0.40 - 0.65 when the altitude is less than 3,500 m. The R-squared values then decrease sharply to 0.20. This indicates the relationship in low elevation area is more significant than in high elevation area. Since frozen surface soil and snow cover occur more often in high elevation area, this result further approves our hypothesis that frozen surface soil and snow cover might disturb the relationships between NCI and observed cloud indices.
| [width=0.8]altitude.eps |
The seasonal variation of R-squared values in the study area is shown in Figure 6. The mean R-squared value at the 7769 pixels is shown by the curve and the standard deviation is shown by error bar. The R-squared values are low in January when it is winter in the Northern Hemisphere. The values become large in March, April and May during the time that the plain area enters spring. The peaks occur in September when the snow and frozen surface soil melt in the Qinghai-Tibet Plateau. The R-squared values drop from October when the snow begins to cover the Qinghai-Tibet Plateau. The standard deviation in summer season is significantly less than it in winter season. This indicates the NDVI is more reliable to estimate cloudiness in summer season. The relatively poor relationship in winter season is probably caused by the snow fall and frozen surface soil.
| [width=0.8]season.eps |
Comparison of R-squared values in January and September is shown in Figure 7. The R-squared values remain high in September for the relationship with both cloud amount and SCI. The lowest values in September occur in high elevation area where the permanently frozen soil and snow cover exist. The R-squared values are lower in January over the whole study area. The lowest values still occur above the altitude of 3,500 m, corresponding to the elevation of the Qinghai-Tibet Plateau. The values in January in the altitude section of 1,000 - 2,000 m, where the Loess Plateau is, decrease to the level in high elevation area in September. The R-squared values keep high in the alluvial plain, where the altitude is less than 500 m.
| [width=0.8]jansep.eps |
Our study indicates that the daily NCI versus observed cloud amount and SCI relationships are strong over most of the well-vegetated land uses. The strong relationships occur in high elevation area in summer season, and the poor relationships occur in low elevation area in winter season. The frozen surface soil and snow cover might bar the NDVI index from determining cloudiness. The application of daily NDVI index may be a useful tool for estimating clouds influence to solar radiation over a large area.