It follows a "sliding window" algorithm with overlapping windows. The amount of forest and its occurence as adjacent forest pixels within fixed- area "moving-windows" surrounding each forest pixel is measured. The window size is user-defined. The result is stored at the location of the center pixel. Thus, a pixel value in the derived map refers to "between-pixel" fragmentation around the corresponding forest location.
As input it requires a binary map with (1) forest and (0) non-forest. Obviously, one can replace forest any other land cover type. If one wants to exclude the influence of a specific land cover type, e.g., water bodies, it should be classified as no-data (NA) in the input map. See e.g., blog post.
Let Pf be the proportion of pixels in the window that are forested. Define Pff (strictly) as the proportion of all adjacent (cardinal directions only) pixel pairs that include at least one forest pixel, for which both pixels are forested. Pff thus (roughly) estimates the conditional probability that, given a pixel of forest, its neighbor is also forest. The classification model then identifies six fragmentation categories as:
interior: Pf = 1.0 patch: Pf < 0.4 transitional: 0.4 ≤ Pf < 0.6 edge: Pf ≥ 0.6 and Pf - Pff < 0 perforated: Pf ≥ 0.6 and Pf - Pff > 0 undetermined: Pf ≥ 0.6 and Pf = Pff
Last changed: $Date$