module 35: shifting-cultivation disturbance from Sims et al. (2025), replacing the Curtis table

[mscrawford]

Important

To do before this leaves draft

• New input data: point cfg$input in config/default.cfg to an input revision preprocessed with Forest loss drivers: fix the Curtis table, add GFW/Sims et al. (2025) as default source pik-piam/mrland#78 (validation data with calcValidTreeCoverLoss: observed tree cover loss by driver (GFW, Sims et al. 2025) pik-piam/mrvalidation#45 and reportingnames: tree cover loss driver classes pik-piam/magpiesets#10). The model does not compile against the current rev4.136, whose f35_forest_lost_share still has the old columns. A test preprocessing with the three PR heads exists (below); the approved revision follows once the three package PRs are merged and released.
• CHANGELOG: add the CHANGELOG.md entry together with the new input revision.

🐦 Description of this PR 🐦

Model side of pik-piam/mrland#78. The shifting-agriculture disturbance in module 35 (s35_forest_damage 1 and 2, default 2) came from a transcription of Curtis et al. (2018) whose Africa row read 3 Mha instead of 39, so the SSA rate was 13x too low. mrland now derives the rate from the Sims et al. (2025) driver map and writes only the column module 35 uses.

• driver_source shrinks to shifting_cultivation, the one column mrland#78 still writes, and presolve.gms reads it. The file loses overall and the four other Curtis columns.
• combined_loss and s35_forest_damage = 3, which summed shifting agriculture and wildfire, are removed: wildfire is already in the LPJmL carbon densities, the option was not documented, and no config uses it.
• f35_forest_disturbance_share removed from input.gms and input/files: loaded since 2020, read nowhere.
• Otherwise descriptions and comments only (c35_forest_damage_end corrected to s35_forest_damage_end in default.cfg and input.gms). No default changes.

Test preprocessing

Regional and validation data preprocessed with the heads of the three package PRs (pik-piam/mrland#78 at 410607a, pik-piam/mrvalidation#45 at d1f454b, pik-piam/magpiesets#10 at c85ad58; revision 4.136, dev flag crawfordPRBUILD), no warnings.

• Regional data against the released rev4.136: f35_forest_lost_share.cs3 holds the one shifting_cultivation column and f35_forest_disturbance_share.cs4 is gone. Two other files differ through mrland commits already on master: f20_processing_balanceflow.cs3 (six decimals instead of four) and the f15_intake_NIN files (extended; read only with s15_exo_diet = 2).
• Validation data: Resources|Tree Cover Loss by driver is new (117 rows); nothing else changes in size.

Effect on a default run

Six SSP2 runs on this branch with develop of 8 October merged (66aa2bc, which includes #942 and #943), identical except the twelve shifting_cultivation values of f35_forest_lost_share:

• Current: the Curtis values of rev4.136,
• Curtis corrected: the Curtis table as mrland#78 corrects it (src = "Curtis"),
• Sims: the values this PR makes the input (src = "GFW"),

each at the default s35_forest_damage = 2 (full rate to 2025, faded out by 2050). All six use the test preprocessing. All time steps locally optimal.

Against historical data

The three default-setting runs against the historical series of the validation data, 1995 to 2030.

🔧 Checklist for PR creator 🔧

• Label: High risk (default results change through the input data)
• Self-review own code (scenario_config.csv: non-applicable, no default changed; magpie4: no change needed)
• Document changes: CHANGELOG.md with the input revision; in-code descriptions updated; goxygen not run
• Perform test runs: the six runs above on the test preprocessing
• Reporting produces no errors and no new warnings (same reporting functions and statuses in all six runs)

📉 Performance 📈

• This PR's default: 32 mins GAMS (as shipped: 36 mins)

[mscrawford]

To @flohump, I'd like to also suggest that we take this opportunity to switch from automatically fading shifting cultivation out in the default run. I could see an argument that under SSP2 urbanization assumptions, e.g., there will be a decrease in subsistence farming, but I wouldn't say that's necessarily more likely than not, and a more parsimonious treatment of the "default" future would be to leave baseline emissions at today's level. I could of course see SSP1 gradually transitioning to have none (though even so, as you can see this fade is abrupt!). But SSP2, much less SSP3, I'm skeptical.

[mscrawford]

Edit, 6 Oct: combined_loss and option 3 are now removed rather than kept with shifting cultivation alone, where they would only duplicate option 1. The reason is unchanged: wildfires and other natural disturbances, the only sources that aren't endogenously modelled in MAgPIE, are included in our LPJmL carbon densities, so turning them on would be double accounting. I'm still interested in eventually integrating fire and drought endogenously into MAgPIE as degradation sources, which would mean either asking LPJmL to run no-fire simulations, or figuring out a solution with the new SPITFIRE LPJmL.

[mscrawford]

Correction: this is f35_forest_disturbance_share (FRA-based disturbance shares, added in December 2020), which was loaded but never read. Michael Windisch's forest shock table (f35_forest_shock.csv, s35_forest_damage = 4) is not touched.

[flohump]

Thanks for catchting the Africa transcription error, and for adding the Sims et al data.
The code looks good to me.

Just one request and a question:

• Could you add validation to the figures (forest, LUC CO2 emissions) so that we can see if we move closer to validation data or away.
• Are your model runs from the current develop?

[mscrawford]

Thanks for catchting the Africa transcription error, and for adding the Sims et al data. The code looks good to me.

Just one request and a question:

* Could you add validation to the figures (forest, LUC CO2 emissions) so that we can see if we move closer to validation data or away.

* Are your model runs from the current develop?

Hey Florian, the newest plots are from a merge with the current develop. I added the validation plots as well. It's a mixed picture -- the LAM values, strictly speaking, move farther away from the validation data (primary forest area, emissions) while the rest improve. The prediction error for Sims was highest in LAM, so to some degree it might be missing shifting cultivation emissions there, arguing for Curtis-corrected. But my feeling is that if the emissions were "better" in LAM before, it was just happy coincidence, and this newer dataset is more robust.

[flohump]

Thanks for the update. The validation figures are helpful. It seems that we are missing permanent deforestation in LAM. This is an issue independent from this PR.

[flohump]

Thanks. The changes looks fine to me.
With the Sims data, the LAM CO2 LUC emissions are below the validation data. This should be investigated independently from this PR.