The problem. Given sites that each hold some species, choose k sites to protect as many species as possible (the maximum coverage problem), or the fewest sites that protect every species (minimum set cover). Both are NP-hard in general, but with 12 sites we can check all 212 = 4096 subsets and so know the true optimum.
Reading the 2D view. The plan map shows the twelve parcels by habitat (forest, grassland, wetland, rock) with the number of species each holds; protected parcels get a thick green fence line. The matrix has one row per site and one column per species (rarest species on the left): a filled cell means the species lives there. The strip above the columns is green when some protected site holds that species and red when it is lost. Rows of two sites that fill the same columns are redundant, which is exactly what complementarity avoids.
Complementarity. Picking the richest sites first often protects the same common species repeatedly. The greedy rule adds, at each step, the site with the most not yet protected species (Kirkpatrick 1983; Margules & Pressey 2000). Greedy is guaranteed to reach at least 1 − 1/e ≈ 63 % of the optimal coverage for any k (Nemhauser et al. 1978) and to need at most about ln n + 1 times the optimal number of sites for full cover; in practice usually much closer.
Random baseline. If k sites are drawn at random, species s found in ns of the n sites is missed with probability C(n−ns, k)/C(n, k), so the expected species count is Σs [1 − C(n−ns, k)/C(n, k)] (hypergeometric).
Simplifications and honesty. The species-by-site table is synthetic (seeded random: habitat specialists, generalists and rare endemics) and is not data for any real region. Real reserve design also weighs land cost and area, connectivity, edge effects, population viability, threats, climate shifts and abundance, none of which are modelled here (tools such as Marxan solve those with simulated annealing). Presence/absence only; all species count equally. The parcel pictures are illustrations of habitat type, not terrain.
Checks. The Monte-Carlo mean of random selections matches the exact hypergeometric formula; the minimum cover size from subset enumeration matches a separate depth-first search over species; greedy never falls below the 1 − 1/e guarantee. New in 2D: coverage recounted straight from the displayed matrix (column by column) matches the bitmask coverage for your selection, the greedy and the best-possible sets, and the matrix row totals sum to the same number as the column totals.