Land-cover transitions, with every input visible.

Two four-by-four classification grids.

Both examples have 16 cells with categories F = forest, C = cropland and U = built-up. They are synthetic categorical arrays without Earth coordinates, satellite dates or assigned land area. All cells have the same illustrative weight, but we do not imply that one cell equals any real ground area.

The reproducible class totals.

Run the supplied JavaScript module or independently count each cell. The first grid has 6 forest, 5 cropland and 5 built-up cells. The second has 4 forest, 3 cropland and 9 built-up cells. The difference is −2 forest, −2 cropland and +4 built-up fictional cells.

Class totals are not sufficient to describe every transition. The source also counts whether a cell remained in the same category or changed between classes, which is often essential for interpreting classification maps.

Why a geographic result is more demanding.

• Georeferencing, geographic extent, cell size and consistent spatial resolution are required for a real-area estimate.
• Image acquisition dates, sensor and seasonal differences, cloud masks and radiometric differences affect comparability.
• Class definitions, reference samples and independent classification accuracy checks are needed.
• Observed differences must not be confused with a hypothetical future land-use scenario or uncalibrated simulation.