Carbon-conversion Take 5

Updating workflow to use standard (current Landsat-agb) predictors and latest Landsat-agb v1.2.0 results.
Author

Lucas Johnson

Published

March 22, 2023

Changes

Previous version

  • SR (Simple ratio), MSR (Modified simple ratio) added as predictors. NSDS removed as a predictor.

  • Landsat-agb predictions updated to v1.2.0

  • Duplicate inventories of plots removed – where multiple inventories for a single plot were present, a random inventory (year) was selected.

  • Litter & Soil elastic-net regression: nested-cv used to produce 20-fold CV accuracy metrics. A standard LOO-CV approach is used to tune and produce a final model.

AGC model

Params

  • n-train: 2005
  • n-test: 502

Call:
lm(formula = formula, data = train_test$training)

Coefficients:
(Intercept)     live_agb  
  0.0008802    0.4999994

Training 1:1 (With plot-level agb estimates as input)

Testing 1:1 (with landsat-agb predictions as input)

Test accuracy:

AGC test-set accuracy. % Improvement relative to a mean model RMSE.
Metric Estimate
RMSE 25.10
% RMSE 41.25
MAE 19.64
% MAE 32.29
ME -2.62
Rsq -1.81
% Improvement 26.72

BGC model

Params

  • n-train: 2005
  • n-test: 502

Call:
lm(formula = formula, data = train_test$training)

Coefficients:
(Intercept)     live_agb  
    0.47929      0.09778

Training 1:1 (With plot-level agb estimates as input)

Testing 1:1 (with landsat-agb predictions as input)

Test accuracy:

BGC test-set accuracy. % Improvement relative to a mean model RMSE.
Metric Estimate
RMSE 4.89
% RMSE 39.53
MAE 3.83
% MAE 30.99
ME -0.53
Rsq 0.47
% Improvement 27.06

Deadwood C model

Params

  • n-train: 210
  • n-test: 53
  • RF hyperparams:
    • mtry: 5
    • min.node.size: 5
    • num.trees.: 500
    • replace: FALSE

Training 1:1 (With landsat-agb estimates as inputs)

Testing 1:1 (With landsat-agb estimates as inputs)

Test accuracy:

Deadwood C test-set accuracy. % Improvement relative to a mean model RMSE.
Metric Estimate
RMSE 6.74
% RMSE 77.65
MAE 5.12
% MAE 58.95
ME 0.24
Rsq 0.09
% Improvement 5.62

Litter C model

Params

  • n-obs: 120
  • Elastic-net hyperparams:
    • alpha: 0.15
    • lambda: 4.1120292

20-Fold-CV 1:1 (With landsat-agb estimates as inputs)

20-Fold-CV accuracy:

Litter C test-set accuracy. % Improvement relative to a mean model RMSE.
Metric Estimate
RMSE 9.27
% RMSE 69.09
MAE 7.21
% MAE 55.15
ME -0.12
Rsq 0.08
% Improvement 13.20

Soil C model

Params

  • n-obs: - n-obs: 72
  • Elastic-net hyperparams:
    • alpha: 0.55
    • lambda: 8.632366

20-Fold-CV 1:1 (With landsat-agb estimates as inputs)

20-Fold-CV test-set accuracy:

Soil C test-set accuracy. % Improvement relative to a mean model RMSE.
Metric Estimate
RMSE 48.97
% RMSE 42.12
MAE 39.12
% MAE 35.12
ME 0.16
Rsq 0.04
% Improvement 32.06