Carbon Conversion - Take 4b: True AGB

Modeling forest carbon pools as a function of aboveground biomass and static climate and topographic predictors.

Changes

This round differs from the last iteration in the following ways:

• True AGB values (measured on plots) used in place of Landsat-derived AGB as a predictor. This document should serve as a comparison/alternative to the previous version.

AGC Model

• \(n\) train: 1838
• \(n\) test: 473
• Predictors: agb, mag, nbr, delta_nbr, ndvi, nsds, delta_nsds, tcb, delta_tcb, tcg, delta_tcg, tcw, delta_tcw, yod, dem, slope, aspect, twi, tmin, tmax, precip, chm, lcpri4, lcpri6, lcpri2, lcpri3, lcsec3, lcsec2, lcsec1, lcsec4

Random Forest Model - Test Set 1:1

Params

$formula
agc ~ .

$num.trees
[1] 1750

$mtry
[1] 20

$min.node.size
[1] 2

$replace
[1] FALSE

$sample.fraction
[1] 0.9

PC Regression - Test Set 1:1

Params

• Number of principle components: 23

Elastic Net Regression - Test set 1:1

Params

• alpha: 1
• lambda: 0.9461285

Test Set Accuracy

	RF	PC	Elastic Net
RMSE	27.609	27.633	27.686
% RMSE	41.029	41.064	41.144
MAE	21.766	21.778	21.843
% MAE	32.345	32.364	32.460
MBE	-3.168	-3.184	-3.092
R2	0.287	0.286	0.283
% Improvement	15.652	15.580	15.416

---

BGC Model

• \(n\) train: 1838
• \(n\) test: 470
• Predictors: agb, mag, nbr, delta_nbr, ndvi, nsds, delta_nsds, tcb, delta_tcb, tcg, delta_tcg, tcw, delta_tcw, yod, dem, slope, aspect, twi, tmin, tmax, precip, chm, lcpri4, lcpri6, lcpri2, lcpri3, lcsec3, lcsec2, lcsec1, lcsec4

Random Forest Model - Test Set 1:1

Params

$formula
bgc ~ .

$num.trees
[1] 450

$mtry
[1] 20

$min.node.size
[1] 2

$replace
[1] FALSE

$sample.fraction
[1] 1

PC Regression - Test Set 1:1

Params

• Number of principle components: 23

Elastic Net Regression - Test set 1:1

Params

• alpha: 0.45
• lambda: 0.017364

Test Set Accuracy

	RF	PC	Elastic Net
RMSE	5.411	5.424	5.427
% RMSE	39.428	39.519	39.542
MAE	4.296	4.294	4.297
% MAE	31.299	31.284	31.306
MBE	-0.578	-0.639	-0.629
R2	0.263	0.260	0.259
% Improvement	14.247	14.049	14.001

---

Deadwood C Model

• 5 filtered
  • Outlier threshold = [(5 * IQR) + 0.75th percentile] = 54.941
• \(n\) train: 374
• \(n\) test: 97
• Predictors: agb, mag, nbr, delta_nbr, ndvi, nsds, delta_nsds, tcb, delta_tcb, tcg, delta_tcg, tcw, delta_tcw, yod, dem, slope, aspect, twi, tmin, tmax, precip, chm, lcpri4, lcpri6, lcsec2, lcsec3, lcsec1, lcsec4

Random Forest Model - Test Set 1:1

Params

$formula
deadwood ~ .

$num.trees
[1] 60

$mtry
[1] 3

$min.node.size
[1] 6

$replace
[1] FALSE

$sample.fraction
[1] 1

PC Regression - Test Set 1:1

Params

• Number of principle components: 22

Elastic Net Regression - Test set 1:1

Params

• alpha: 0.4
• lambda: 0.2570002

Test Set Accuracy

	RF	PC	Elastic Net
RMSE	5.866	6.573	6.321
% RMSE	77.653	87.005	83.679
MAE	4.488	5.004	4.697
% MAE	59.413	66.241	62.177
MBE	0.909	0.747	0.018
R2	0.182	-0.026	0.051
% Improvement	10.043	-0.791	3.061

---

Litter C models

• \(n\) train: 90
• \(n\) test: 28
• Predictors: agb, mag, nbr, delta_nbr, ndvi, nsds, delta_nsds, tcb, delta_tcb, tcg, delta_tcg, tcw, delta_tcw, yod, dem, slope, aspect, twi, tmin, tmax, precip, chm, lcpri4, lcsec3, lcsec1

PC Regression - Test Set 1:1

Params

• Number of principle components: 19

Elastic Net Regression - Test Set 1:1

Params

• alpha: 0.15
• lambda: 4.3322608

Test-set Accuracy

	PC	Elastic Net
RMSE	14.389	8.428
% RMSE	84.827	49.682
MAE	9.571	6.731
% MAE	56.421	39.682
MBE	-6.848	0.000
R2	-0.377	0.528
% Improvement	-15.238	32.507

---

Soil C model

• Predictors: agb, mag, nbr, delta_nbr, ndvi, nsds, delta_nsds, tcb, delta_tcb, tcg, delta_tcg, tcw, delta_tcw, yod, dem, slope, aspect, twi, tmin, tmax, precip, chm, lcpri4, lcsec3
• \(n\) train: 53
• \(n\) test: 17

PC Regression - Test Set 1:1

Params

• Number of principle components: 18

Elastic Net Regression - Test Set 1:1

Params

• alpha: 1
• lambda: 7.609536

Test-set Accuracy

	PC	Elastic Net
RMSE	49.386	42.406
% RMSE	44.033	37.810
MAE	37.958	30.282
% MAE	33.843	26.999
MBE	1.879	0.000
R2	0.311	0.492
% Improvement	19.476	30.856