Changes since 1.0.0
Predictors
Landsat-derived indices
The following predictors were added:
- NDVI (Normalized difference vegetation index)
- SR (Simple ratio)
- MSR (Modified simple ratio)
LT-GEE segmentation parameters were modified for smoothing and gap-filling all spectral-indices (TC*, NBR, NDVI, SR, MSR, Delta_*) to be less sensitive. Parameters for disturbance metrics computed with NBR (Year-of-disturbance, Magnitude-of-disturbance) were left unchanged.
| Parameters | Spectral (NBR, NDVI, SR, MSR, TC*) | Disturbance (YOD, MAG) |
|---|---|---|
| maxSegments | 5 | 10 |
| spikeThreshold | 0.5 | 0.9 |
| vertexCountOvershoot | 3 | 3 |
| preventOneYearRecovery | true | true |
| recoveryThreshold | 0.25 | 0.75 |
| pvalThreshold | 0.05 | 0.05 |
| bestModelProportion | 0.75 | 0.75 |
| minObservationsNeeded | 6 | 6 |
| Parameter | Value | Operator |
|---|---|---|
| Delta | Loss | |
| Sort | Most recent | |
| Year | 1985-Target year | |
| Magnitude | 50 | Greater than |
| Duration | 4 | Less than |
| Pre-disturbance spectral value | 300 | Greater than |
| Minimum mapping unit | 7 |
Ecological data
The following predictors were added after this analysis:
dist_to_water: distance in meters to census area/line water (downloaded from tigris)ecozones: Epa level 4 ecozones. When level 4 ecozones did not cover>=2% of the state area, i aggregated to level 3 ecozone. If that aggregation did not yield>=2% of the state area, I set toother.wetlands: FWS wetland classifications
Direct Modeling
All FIA plots were made available (2002-2019) rather than the most recent cycle (2013-2019)
FIA plots were partitioned such that a single panel was randomly selected for map assessment, and the remaining 4 panels were left for model development.
LiDAR-zeroes: Plots labeled completely non-forest by FIA and with LiDAR-derived max-heights <= 1 included in model training.
I simplified the linear ensemble models to remove interactions between terms.
Indirect Modeling
- I switched our v1.4 lidar-agb predictions to the predictions developed in the lidar-agb manuscript. a) this made the explanation in the landsat-agb manuscript easier – now I can simply refer to that paper without any caveats. b) The max AGB predictions were higher for the manuscript models (~320 vs ~280) and given that we were struggling with saturation, I thought this might help.
- I limited the sample to only grass/shrub, wetland, and tree cover LMCAP classes. Previously we had been including cropland in the sample, but excluding cropland later in the final mask for landsat-agb. So I thought excluding cropland would give the indirect model a better fighting chance – only train on what we will eventually assess.
- I simplified the linear ensemble models to remove interactions between terms.
Ensemble
- The ensemble is now a simple average:
[(direct pred + indirect pred) / 2]