Skip to content

Validation — disslucc vs TerraME

The agreement between disslucc and the original TerraME/LuccME is checked in a separate repository, LambdaGeo/disslucc-benchmark, which keeps the reference results, the comparison code and the numbers. This repository holds the package and examples only, so that learning the model and checking it against TerraME are two separate things.

Where things are

What Where
Per-lab results (iterations, MAE, max error, timings), reproducible with make benchmark disslucc-benchmark
TerraME/LuccME reference results (goldens, all 21 LuccME labs) and the container that generated them LambdaGeo/luccme-goldens, terrame-docker
What is checked here tests/: demand components, and the TOML executors against hand-built experiments (same output file, same SHA-256)

disslucc-benchmark pins the disslucc release it was run with; the numbers it reports belong to that pair of versions.

Two scenarios worth knowing

The package's two original scenarios are kept in the benchmark under these ids:

  • lab01_md1643 (continuous, Amazon deforestation, csAC region, 6,574 cells): iteration counts identical to TerraME in every year.
  • lab15_md10 (discrete, Moju region, 5,914 cells): iteration counts identical to TerraME in every year.

Their numbers are in the benchmark report; they are not copied here, so there is one place to update when they change.

One deliberate difference

LuccME's correctCellChange never runs: its guard reads cell.regionregionAloc, a field name that no cell has (in AllocationCClueLike.lua). disslucc runs the correction by default (AllocationClueLike(cell_correction=True)), the algorithm as documented. With cell_correction=False disslucc follows the behavior of the reference run. The benchmark reports both, and labels the difference "by design".

History

Before 0.5.0 this repository carried its own goldens, a Lua differential test and a discriminance suite. They moved to (or were superseded by) the benchmark. The full set is in the v0.4.0 tag (browse), including the earlier version of this page with the Lab 1 and Lab 15 analysis.