Public repository for the arXiv-ready preprint:
Transparent pre-training diagnostics for Maskable-PPO land-use decision support on irregular parcels
The paper asks whether a Maskable-PPO scorer-MLP policy adds practical decision-support value over transparent exact-delta local diagnostics for cadastral-style farmland--forest exchange. The conclusion is deliberately bounded: in the tested low-intervention, sparse-adjacency synthetic parcel regimes, exact-delta local diagnostics were stronger and much cheaper than the tested DRL family.
The current public manuscript entry point is:
| Path | Contents |
|---|---|
manuscript/main.tex |
Public arXiv preprint source |
manuscript/main.pdf |
Compiled public preprint PDF |
manuscript/figures/ |
Figures used by the public preprint |
submission/eswa_anonymous/ |
Historical ESWA transfer package retained for provenance, not the current public entry point |
- A slope-exact local baseline reduces area-weighted farmland slope by 5.9 to 9.7 percentage points across three synthetic regions.
- A contiguity-aware exact-delta local sweep recovers slope-reducing non-negative-contiguity points in every region.
- The sampled local grid dominates the six tested DRL mean outcomes and most individual deterministic rollouts.
- Dense scalarisation, no-reuse parity, two-step lookahead, fixed-policy budget decoding, matched Region B retraining, stochastic inference, neighbour-overlap analysis, and CV(area)--budget stress tests support the same screening conclusion.
- The tested policies learn spatial clustering, but weak effective inter-swap coupling and train--evaluation mismatch limit deterministic decision-support value.
| Path | Contents |
|---|---|
manuscript/ |
Public preprint PDF, source, and figure files |
experiments/ |
Audit scripts, audit notes, and CSV/JSON outputs used by the manuscript |
synthetic_data_gen/ |
Public-data synthetic parcel generation pipeline |
v1_compat/ |
Synthetic regions in CSV plus sparse adjacency format |
v1_scripts/ |
DRL training, evaluation, baselines, and environment code |
submission/eswa_anonymous/ |
Archived ESWA submission package for provenance |
The public-data pipeline can regenerate the synthetic parcel structures:
pip install -r synthetic_data_gen/requirements.txt
cd synthetic_data_gen
python run_all.pyEarth Engine authentication and a configured Google Earth Engine project are required. The pipeline writes parcel polygons, parcel features, and Queen adjacency for the three synthetic regions.
The experiments/audits/ and experiments/results/ directories contain the diagnostic evidence behind the manuscript:
- contiguity-aware local sweeps
- dense local-frontier audit
- no-reuse local parity audit
- two-step local lookahead audit
- budget-sensitivity and fixed-policy decoding audits
- CV(area) and joint CV(area)--budget stress audits
- structural-validity and graph-contiguity checks
- large-budget GA audit
The synthetic data generation pipeline, trained model artefacts, paired-inference evaluation outputs, baseline results, audit outputs, figure-generation records, and figures are released in this public repository. Historical anonymous review links in earlier submission materials are superseded by this public GitHub repository:
https://github.com/zhouning/cadastral-drl-synthetic
The evidence supports a bounded conclusion about the tested Maskable PPO + scorer-MLP family under low-intervention synthetic cadastral regimes. It does not rule out graph encoders, pair-scoring policies, attention models, inference-time search, stronger tuned metaheuristics, or real cadastral settings with stronger inter-swap coupling.