*4.2. Delineating the Endmembers Manually*

Accurate and fast endmember selection is the key to the successful application of the method. Although the endmember selection based on high-resolution remote sensing images (named ESRS) could ensure the purity and representativeness of the endmembers, it suffered from the heavy load of work for pure pixel selection and was highly dependent on the availability of high-resolution images.

Therefore, we explored the feasibility of manually delineating endmembers (named MDE). This was achieved by drafting the vertex of the convex simplex through human– computer interaction in the feature space. We first delineated those pixels around the vertices as endmembers, as close to the vertices as possible. As this work was susceptible to the operator's knowledge of what a "vertex" is, the operation was repeated three times by covering different area sizes (Figure 13).

**Figure 13.** Delineating endmembers with different area sizes: size 1 (**a**), size 2 (**b**), size 3 (**c**). The sizes of the area gradually expand.

The estimated cropping intensity with different vertex sizes was compared with MCD12Q2 at the county level. The R2 of the correlation between the estimation and MCD12Q2 were all above 0.87 (Figure 14), which confirmed that the proposed endmember selection method is applicable and robust. Compared with ESRS, MDE has the advantages of high efficiency and ease of implementation, and the accuracy and efficiency can be well-balanced.

**Figure 14.** Correlation between MCD12Q2 and MDE estimated at the county level.

The accuracy of the unmixing decreased slightly when the sizes of endmember areas expanded (Figure 14). This is explainable since the closer the endmember is to the vertex of the convex simplex, the purer the endmember will be. Therefore, to ensure the accuracy, the delineating area of endmembers should be as small as possible and as close to the vertex as possible.
