*4.3. Unmixing in Regions with Different Sizes and Varied Endmember Land-Cover Types*

The shape of the feature space varied when the research area covered different regions with different sizes. The shape of the feature space affected the unmixing accuracy since the approximate triangle was the basis of the method. Therefore, the relationships of the size of the research area and the completeness of the land-cover types with the unmixing accuracies were explored.

The estimated cropping intensity and MCD12Q2 were compared for test areas with different sizes (Figure 15). The mean cropping intensity values were compared at 10 km × 10 km block level. The area sizes, the corresponding feature space and the unmixing accuracies for each test area are given in Table 3.

The completeness of the three endmembers in the study area was the precondition for the successful application of this method. The three vertices in the feature spaces were obvious in all test areas with varied sizes. R2 values were above 0.87 in all test areas, demonstrating that the size of the study area had little effect on the unmixing accuracy as long as the study area had all necessary land-cover types.

The effectiveness of the proposed method also depends on the completeness of endmember land-cover types. Our method can be applied directedly to the North China Plain and the middle and lower reaches of the Yangtze River Valley, which are double-cropping or double-single-mixed cropping areas with relatively large patches of croplands. The successful application of the method requires the concurrence of three land-cover types (double-cropping, natural vegetation and water bodies). The method will need further work (for example, to construct a new feature space, to find the new optimal endmembers again) when lacking any of the three necessary endmember land-cover types, such as those areas where crops have only one growing season (Table 4). However, expanding the test area does aid the inclusion of all three necessary endmember land-cover types.

**Figure 15.** Test areas with different sizes.




**Table 4.** Completeness of endmembers when test areas are located in regions with varied land-cover types.

More effort is needed to find the proper endmembers when the method is extended to areas other than China. Since the interpretation for PCA images is scene-dependent and there has been no well outlined procedure for it, the interpretation needs more elaboration, and that is where innovation is possible.

The abundance of natural vegetation and water bodies can also be estimated as a "by-product" of this research. MODIS MOD13Q1 products within a year were used and the focus was cropping intensity estimation in this study. If the method is transplanted to time series vegetation indices with different temporal spans, other land-cover types can also be unmixed.
