Landsat FIA AGB 1.1.3: Stratified Sample

Evaluation Results

Last version of FIA models: 2022-06-23 LiDAR Zeroes

Change Summary

• Training data is a stratified sample of FIA plots. We selected an equal number of plots from 25 Mg/Ha bins up to 300 Mg/Ha, and from a final bin > 300 Mg/Ha. The stratified sample size is as close as posssible to the original sample size.

• To fill each bin random sampling with replacement was conducted. In some bins, this results in downsampling, while in other bins this results in upsampling.

	RF (ranger)	GBM (LightGBM)	SVM (kernlab)	Ensemble (model weighted)	Ensemble (RMSE weighted)
%RMSE	42.156	45.158	43.564	41.364	42.158
RMSE	53.275	57.070	55.054	52.275	53.277
MAE	41.092	44.222	42.170	39.599	41.189
MBE	7.223	9.263	6.509	0.986	7.657
R2	0.509	0.456	0.481	0.518	0.509

AGB Distribution

summary(bind_rows(training, testing)$agb_mgha)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
   0.00   80.21  147.76  151.75  223.01  425.00 

Metadata

Ensembles

• RMSE-weighted model weights:

• Linear model weights:

Call:
lm(formula = agb_mgha ~ rf_pred * lgb_pred * svm_pred, data = k_fold_preds)

Residuals:
     Min       1Q   Median       3Q      Max 
-132.418   -0.437    2.563    4.576  219.959 

Coefficients:
                            Estimate Std. Error t value Pr(>|t|)    
(Intercept)               -2.800e-01  3.054e+00  -0.092  0.92695    
rf_pred                   -4.742e-02  1.306e-01  -0.363  0.71657    
lgb_pred                   5.925e-01  1.783e-01   3.324  0.00091 ***
svm_pred                   3.805e-01  1.634e-01   2.329  0.02001 *  
rf_pred:lgb_pred           8.164e-04  1.132e-03   0.721  0.47102    
rf_pred:svm_pred           3.642e-03  1.209e-03   3.013  0.00263 ** 
lgb_pred:svm_pred         -4.119e-03  8.824e-04  -4.668 3.31e-06 ***
rf_pred:lgb_pred:svm_pred -3.010e-07  8.371e-07  -0.360  0.71923    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 31.06 on 1522 degrees of freedom
Multiple R-squared:  0.8948,    Adjusted R-squared:  0.8944 
F-statistic:  1850 on 7 and 1522 DF,  p-value: < 2.2e-16

Coverages

• 0 coverages:

\(n\) and \(p\)

• 2196 observations
  • 1534 training
  • 662 testing
• 162 predictors
  • aweinsh, delta_aweinsh, aweish, delta_aweish, bcc, delta_bcc, dvi, delta_dvi, embi, delta_embi, exgr, delta_exgr, exr, delta_exr, fcvi, delta_fcvi, gari, delta_gari, gbndvi, delta_gbndvi, gemi, delta_gemi, gli, delta_gli, gndvi, delta_gndvi, gosavi, delta_gosavi, grndvi, delta_grndvi, gvmi, delta_gvmi, ikaw, delta_ikaw, ipvi, delta_ipvi, lswi, delta_lswi, mag, mbi, delta_mbi, mcari2, delta_mcari2, mlswi26, delta_mlswi26, mlswi27, delta_mlswi27, mndwi, delta_mndwi, mrbvi, delta_mrbvi, msavi, delta_msavi, msi, delta_msi, msr, delta_msr, mtvi2, delta_mtvi2, muwir, delta_muwir, nbr, delta_nbr, nbr2, delta_nbr2, nbsims, delta_nbsims, nddi, delta_nddi, ndsi, delta_ndsi, ndvi, delta_ndvi, ndvimndwi, delta_ndvimndwi, ndwi, delta_ndwi, ngrdi, delta_ngrdi, nirv, delta_nirv, nli, delta_nli, nmdi, delta_nmdi, normg, delta_normg, normnir, delta_normnir, normr, delta_normr, nrfig, delta_nrfig, nsds, delta_nsds, nwi, delta_nwi, pisi, delta_pisi, rcc, delta_rcc, rdvi, delta_rdvi, rgbvi, delta_rgbvi, rgri, delta_rgri, s3, delta_s3, si, delta_si, sr, delta_sr, sr2, delta_sr2, swi, delta_swi, tcb, delta_tcb, tcg, delta_tcg, tcw, delta_tcw, tdvi, delta_tdvi, tgi, delta_tgi, trivi, delta_trivi, tvi, delta_tvi, ui, delta_ui, vgnirbi, delta_vgnirbi, vig, delta_vig, vrnirbi, delta_vrnirbi, wi1, delta_wi1, wi2, delta_wi2, yod, precip, tmin, tmax, dem, twi, slope, aspect, chm, lcpri4, lcpri6, lcpri3, lcpri2, lcpri5, lcsec3, lcsec2, lcsec1, lcsec4, lcsec8

Component Models

• Tuning used 5-fold CV
• Final hyperparameters:

Random forest:

$num.trees
[1] 2000

$mtry
[1] 16

$min.node.size
[1] 2

$sample.fraction
[1] 1

$replace
[1] FALSE

$formula
agb_mgha ~ .

LGB:

$nrounds
[1] 500

$params
$params$learning_rate
[1] 0.1

$params$num_leaves
[1] 59

$params$max_depth
[1] 19

$params$extra_trees
[1] TRUE

$params$min_data_in_leaf
[1] 14

$params$bagging_fraction
[1] 0.6

$params$bagging_freq
[1] 1

$params$feature_fraction
[1] 0.5

$params$min_data_in_bin
[1] 18

$params$lambda_l1
[1] 0.5

$params$lambda_l2
[1] 0.1

$params$force_col_wise
[1] TRUE

SVM:

$x
agb_mgha ~ .

$kernel
[1] "laplacedot"

$type
[1] "eps-svr"

$kpar
$kpar$sigma
[1] 0.00390625


$C
[1] 84

$epsilon
[1] 0.00390625

$nu
[1] 0.2