IVES ET AL. · PARTS PRINCIPLES
Dependence changes the uncertainty.
This experiment uses actual acquisition gaps, cell-level GLS, and cross-partition covariance. It is a continuous-time adaptation of PARTS principles, not a run of the native remotePARTS package.
Research intervals only. Re-estimated-scale partition intervals failed the conditional simulation check: about 76% coverage in the primary scenario, despite a nominal 95% level. No experiment interval is released as calibrated forest-damage confidence.
Fixed scope: 176 cells around 44.871366, −88.871979, 600 m radius. These regional results do not change when the map coordinate changes. They are not cell-level probabilities or estimates for the whole project.
The equal-cell mean estimates the mean fitted trend on this finite map. Full and partition GLS use different weights and target a common-mean working model. The dashed diagnostic omits cross-partition covariance; it is not a valid independence assumption. Units: index yr⁻¹. All intervals use ±1.96 × plug-in standard error.
Scenario diagnostics
Known-null coverage check
Each registered sentinel has 500 simulated repetitions per index. “Oracle” knows the generating covariance; “refit” re-estimates residual scales and model coefficients, while covariance shapes remain fixed. Monte Carlo limits describe simulation precision, not forest accuracy. Other selected scenarios were not simulation-tested.
Follow-up: include temporally independent observation noise
After the first failure, an additive sensitivity experiment tested 25% and 50% temporal nugget. The initial grid and failed results remain intact. This is exploratory follow-up, not a newly selected production model.
A temporal nugget here is independent across dates but retains the model's spatial covariance. It does not prove sensor noise is spatially independent.
FOLLOW-UP · INDEPENDENT OUTER SIMULATION
Can recalibrating the interval help?
Parametric bootstrap-t refits residual scales and partition weights in each simulation. Every outer dataset runs its own 399-draw inner bootstrap. The 200 outer datasets assess coverage independently of the inner calibration draws.
This fixed experiment is NDRE only: 30-day temporal scale, 150 m spatial range, 25% shared component. It targets a common slope, not the finite-map average. The second result remains below 95%; neither result establishes real-data coverage. The original failed method and all scenarios are retained.
Nested simulation, interval width & hashesOBSERVATION AND MODEL SENSITIVITY
Inspect each local candidate's evidence.
The 11-cell union contains the six original candidates plus cells flagged in source-choice scenarios. Counts measure deliberate sensitivity checks, not repeated independent trials.
Spatial designs, deleted years, and source-choice scenarios reuse data. Combining their pass rates cannot produce a damage probability. Cells outside this 11-cell audit are not assigned a borrowed result.
Download all local evidence axes · CSVMONITORING AND EXTERNAL VALIDATION
What has actually been tested?
The 2025 comparison has only two dates and uses a retrospective sampling frame. It is a conditional pseudo-time-out test. A more complex model did not consistently beat the previous-year median.
Reproduce and inspect the evidence.
The existing map and its candidate rules remain v4. This page adds evidence and research experiments; it does not relabel sensitive candidates as false positives.