Isolation Forest
Nine variants: sample settings 64, 128 and 256, with three seeds each. Shorter corrected paths produce larger raw anomaly scores.
LIVING ATLAS / METHODS
Methods.
Three model families.
One fixed regional trend protocol.
Follow one workflow.
Quality-controlled imagery → CELL observations → anomaly screening → support and decline checks → regional trend interval.
The three anomaly detectors and original review candidates are preserved. The final analysis follows the fixed parts-core-equal-mean-1.0 protocol. Anomaly screening identifies unusual changes; the regional interval describes uncertainty in the average trend of the fixed study area.
Mahalanobis distance
Three diagonal-shrinkage settings: 0.10, 0.30 and 0.60. The distance describes how far a change pattern lies from the training reference.
PCA reconstruction
Three training-variance targets: 0.80, 0.90 and 0.95. Root-mean-square reconstruction error uses all eight input features.
Each fitted variant is calibrated against its own held-out reference. Average ranks within a family first, then average the three family means with weights of one third. Directly averaging all 15 variants would incorrectly give Isolation Forest 60% of the weight.
One fixed area. One trend interval.
The completed 600 m corridor at 44.871366, −88.871979 contains 176 CELLs, 21 usable dates and 3,480 CELL-date observations. NDRE and NDVI use the same missing-data mask.
Fit a temporal GLS trend to each CELL, then average all 176 slopes equally. Propagate the complete covariance between CELLs, including its off-diagonal entries, to obtain the regional standard error and nominal 95% normal interval. Anomaly scores are not GLS weights, and the analysis is not restricted to candidate cells.
| Fixed setting | Value |
|---|---|
| Temporal correlation scale | 30 days |
| Spatial correlation scale | 150 m |
| Shared spatial component / spatial nugget | 0.25 / 0.25 |
| Temporal nugget | 0 |
These are fixed working assumptions from the existing primary scenario, not fitted optimal values. This workflow does not partition the region, search parameter settings, or run a nested bootstrap.
The interval applies only to this frozen area. It is not an interval for an individual CELL, a damage probability, or a future forecast. A different location or input requires a new versioned protocol. Both current regional intervals include zero; that does not rule out local change.
Fixed protocol · JSON · Regional summary · CSV · CELL details · CSV
Read each result at its own scale.
Each CELL retains its three family ranks, ensemble score, agreement count, and original candidate status. Available dates, pixel coverage, GLS slopes, and saved source and year-deletion checks are read by exact ID. A missing result stays missing; another CELL is never substituted.
Source checks and year-deletion checks reuse the same observations. Their counts cannot be multiplied or averaged into a confidence percentage. Spatial sensitivity details remain in downloads and the research archive.
The covariance parameters are treated as fixed. Their selection uncertainty is not included. Retrospectively selected controls are not confirmed healthy labels, and independent ecological accuracy still needs field observations or independent labels.
Inputs & reproducibility.
Image quality, common forest, and grid definitions
99 PlanetScope 8-band SR / UDM2 scene groups cover 26 July dates from 2021–2025. Accept clear pixels with confidence ≥50 and positive bands and denominators. Same-day overlaps keep the highest confidence; ties use the first scene ID in lexical order. Missing observations are neither zero-filled nor interpolated.
NDRE = (B8 − B7) / (B8 + B7); NDVI = (B8 − B6) / (B8 + B6). Index values outside [−1, 1] are missing. Forest context is 5 ≤ historical CHM < 100 m, with common-valid forest observed in each of the five years.
The project samples 75 m CELLs at 105 m spacing. The local refinement uses globally aligned, complete 75 m CELLs whose centers fall within 600 m. These grids and their coverage rules are not interchangeable. The fixed core cohort has 19–21 usable dates per CELL; 2025 contributes only two dates.
Original training, candidate rules, and validation
Training-only median/MAD standardization uses eight annual differences from 2021. Five hashed groups of 1.5 km blocks separate training, calibration, and query roles with a 150 m reference buffer.
Local candidates require ensemble ≥0.95, at least two family ranks ≥0.95, reference distance >150 m, at least two median valid pixel dates per year, NDRE Huber slope ≤−0.002 per year, and 2025−2021 NDRE ≤−0.01. Project candidates instead use annual Theil–Sen slope ≤−0.005, change ≤−0.025, and at least three declining annual steps. These rules remain separate and unchanged.
Loading original validation results…
Original validation · JSON · Original reproducibility manifest · JSON
Research archive and source checks
Earlier parameter scenarios, dependence experiments, temporal contrasts, partitions, and bootstrap work remain available as research records. They are not additional steps in the current primary workflow.
Advanced calculation archive · Spatial and year sensitivity · CSV · Saved source sensitivity · CSV
Original imagery and historical results remain preserved. Current formulas, conventions, frozen input hashes and acceptance values are documented in the current PDF and fixed protocol.