Calculation lab.
Observed changes. Three model families. One equally weighted score.
Loading the original fitted models…
Inputs → models → fusion.
The original model scores and review candidates stay unchanged. A higher score means a more unusual spectral trajectory.
Select a cell to inspect
1. Observed changes
Eight inputs: NDRE and NDVI in 2022–2025, each minus its 2021 value. Training controls provide the median and MAD used for standardization.
Inspect the eight inputs and training scales
Scale = max(1.4826 × MAD, 0.00001). Training, calibration and query roles remain separated.
2. Three model families
The 9 + 3 + 3 fitted variants are retained. Each variant is ranked against its own held-out calibration reference.
Inspect a fitted model and its calibration
3. Equal-family fusion
Average calibrated ranks inside each family, then average the three family means. Each family contributes one third.
Directly averaging all 15 variants would give Isolation Forest 60% of the weight. This score is not a damage probability or a conformal p-value.
4. Support & saved checks
Dates, coverage, and source / year sensitivity
Saved checks reuse the same data. Their votes are not independent and are not combined into a confidence percentage. Results are matched by exact CELL ID.
Open this cell’s evidence →The regional trend is a separate result.
The fixed 176-cell corridor uses temporal GLS and the full covariance between cells. Its equal-cell regional interval is available in Evidence; changing a map coordinate does not transfer that interval to a new area.
Open fixed-area Evidence · Advanced research archive · Methods