*4.1. Determining the Final Endmembers*

To demonstrate the validity of the final endmembers, three PCA components were used to visually examine the distribution of the candidate endmembers in the feature space (Figure 12). Through rotating the feature space, different PCA component combinations could be obtained and be represented in this three-dimensional space. For the two-dimensional case, it was viewed by rotating the feature space to compress one of the dimensions. It was difficult to visualize the one-dimensional case, so it was carried out in the two-dimensional space and assumed to be projected into one of the features. PCA 2 was found to be the most ideal feature when one feature was used. Table 2 lists all possible combinations using the three PCA components.

Through the visual examination, we found that three models might meet our requirements (Table 2, in bold font). With the first model, it was difficult to meet the purity requirements for the endmembers because it lacked major land-cover types. The fourendmember model (the fifth model) did not fit the definition of cropping intensity in our study because single-cropping was also taken as one of the endmembers. Therefore, the three-endmember model using PCA 1 and PCA 2 was finally selected.

**Figure 12.** The three-dimensional feature space.

**Table 2.** Multiple endmember models.

