How to build a land-cover transition matrix for a national report

A transition matrix cross-tabulates baseline class against target class, area by area. Rows are the class at your baseline date, columns are the class at your target date, and each cell holds the area that moved from one to the other. The diagonal is what stayed put. Everything off the diagonal is the change story: forest to cropland, wetland to built-up, grassland to bare soil. If you report against SDG 15.1.1 or feed a UN-SEEA land account, a reviewer will usually ask for this table before they look at the headline percentage.
Building one sounds mechanical. In practice most of the work happens before you ever cross-tabulate anything.
Get the inputs into a state where cross-tabulation means something
The matrix is only as good as the two layers going into it. Three things to check before you run anything:
- Same legend, both dates. If your baseline classification used a different class scheme than your target year, your cells don't mean what you think they mean. A merge from six classes to four between cycles will show "change" that's actually just a reclassification.
- Same pixel grid. Co-registration matters more than people expect. A few pixels of spatial offset between two independently produced layers shows up as spurious transitions all along every class boundary, and boundary pixels are a large share of total area in most landscapes.
- Same minimum mapping unit. Mixing a fine mapping unit at baseline with a coarser one at the target year inflates transitions in fragmented areas purely from the generalization difference, with no real change on the ground behind the numbers.
This is the part that burns a reporting cycle. Reclassifying an old study to match a new legend, or re-deriving a prior year's layer because the methodology shifted between commissioned studies, is slower than the actual cross-tab.
Running the cross-tab and getting from pixels to a defensible area figure
Once the two rasters are aligned and share a legend, the cross-tabulation itself is a GIS operation: intersect the two classified layers, count pixels (or area) per combination, and populate the matrix. The raw pixel counts give you a first pass, but pixel counts alone are not the number that goes in a national report. Two adjustments usually follow:
- Convert to area using the pixel resolution, and keep the unit consistent with how your statistics office reports elsewhere (hectares for land accounts is standard).
- Apply an accuracy-based area estimator if you're reporting a figure with any statistical weight behind it. The stratified estimator approach (Olofsson et al. is the usual citation) corrects the naive pixel-count area using a confusion matrix from reference samples, and gives you a confidence interval to publish alongside the figure.
The output table you hand to the report writer is the matrix itself plus net change by class (gains minus losses) and, where it's asked for, gross change (gains plus losses treated separately, since a landscape can have a lot of churn with very little net change).
Where this breaks down cycle after cycle
For a statistics team, the recurring problem is that baseline and target layers for the matrix often come from separately commissioned classification work, each with its own legend, its own mapping unit, and its own analyst's judgment calls. Every cycle a new team has to reconcile that against what came before, and the transition matrix for the current period is only as defensible as the agreement between two classifications that were never designed to be compared.
That's the gap a consistent, annually-refreshed land-cover change layer is meant to close: one classification scheme carried forward year over year, so the matrix you build this cycle sits on the same legend and grid as the one you built last cycle, instead of starting the reconciliation work over.
If your transition matrix keeps breaking at the "are these two years even comparable" step, that's the problem Habitat Loss Mapping is built to remove.