Post-classification comparison vs direct multidate classification
Every national land-cover change figure eventually gets this question from someone on the SDG reporting side: did you classify each date separately and compare the maps, or did you run one classification on a stacked multidate image? The two approaches are called post-classification comparison and direct multidate classification, and the choice between them shows up in your error matrix, your audit trail, and the explanation you give when a province's forest-loss number jumps between reporting cycles.
What each method actually does
Post-classification comparison classifies each date independently, then compares the two label maps pixel by pixel. If a pixel was "forest" in 2015 and "cropland" in 2020, that's a change record. Direct multidate classification skips the separate classification step. It stacks the bands from both dates, or more, into one image and trains a single classifier to recognize transition classes directly, so "forest-to-cropland" is its own class from the start rather than a derived comparison.
Image differencing is a related but simpler idea: subtract band values or an index like NDVI between two dates and threshold the result to flag change. It's fast and needs no training data for the change step itself, but it tells you that something changed, not what it changed into, which matters when your reporting template asks for a transition matrix, not just a change mask.
Where each one breaks down for national reporting
Post-classification comparison has one well-known problem: errors compound. If each date's classification is off by a few percentage points, independently and in different directions, the change map inherits both sets of errors. A pixel that gets flipped between "grassland" and "sparse woodland" across the two classifiers, with no actual change on the ground, shows up as loss or gain in your statistics. For a territory with a lot of transitional or mixed-canopy cover, this can inflate change area in ways that are hard to explain to an auditor.
Direct multidate classification sidesteps that particular failure mode, since there's no independent classification on each date to disagree with itself. It trades that problem for a different one: training data. You need labeled samples for every transition class you care about, and a legend with twelve cover types can turn into dozens of transition classes once you cross them against each other. Training data for rare transitions, say wetland-to-built-up, is often thin on the ground, literally.
In practice, a lot of national programs land on a mixed approach: a cleaned post-classification comparison for the headline statistic, with direct multidate or image-differencing checks used to flag suspicious pixels for manual review before the figure goes into the report. That hybrid is slower to set up but easier to defend when someone asks why a specific hectare count moved.
What this means for your annual figure
None of this is really about picking the theoretically superior algorithm. It's about which method you can reproduce and rerun next year with the same inputs, and get a number that tracks real change rather than classifier noise. A method that scores better in a benchmark paper isn't worth much if nobody on your team can explain, two reporting cycles later, why last year's loss figure and this year's don't reconcile.
That's the practical argument for a change layer built the same way every year on the same imagery source, rather than re-deriving the method from scratch each cycle. It's also the whole premise behind an annually-refreshed land-cover change layer built for exactly this kind of multi-year reporting.
If your next SDG submission needs a change figure you can walk an auditor through line by line, that consistency is worth more than chasing the newest classification technique.