Landsat FIA AGB 1.1.2: PCA (less data = more faster

Evaluation Results

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

Change Summary

• This version uses PCA to reduce the dimensionality of the data

• We used 99% variance to select PCs

• This version includes 176 completely non-forested FIA plots, with maximum LiDAR heights < 1m, and an inferred AGB value of 0 Mg/ha

	RF (ranger)	GBM (LightGBM)	SVM (kernlab)	Ensemble (model weighted)	Ensemble (RMSE weighted)
%RMSE	44.685	43.180	41.051	41.236	41.963
RMSE	56.472	54.569	51.878	52.112	53.031
MAE	42.660	41.828	39.651	39.606	40.560
MBE	0.674	1.140	-1.150	-0.008	0.214
R2	0.439	0.474	0.526	0.521	0.508

AGB Distribution

summary(bind_rows(training, testing)$agb_mgha)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
   0.00   75.17  126.45  125.90  174.95  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 
-144.838  -33.717   -1.877   28.339  218.126 

Coefficients:
                            Estimate Std. Error t value Pr(>|t|)    
(Intercept)               -5.889e+00  6.757e+00  -0.872 0.383576    
rf_pred                    2.099e-01  1.728e-01   1.215 0.224516    
lgb_pred                  -4.562e-01  2.827e-01  -1.614 0.106808    
svm_pred                   5.407e-01  2.336e-01   2.315 0.020740 *  
rf_pred:lgb_pred           5.451e-03  1.859e-03   2.932 0.003416 ** 
rf_pred:svm_pred           3.690e-04  2.047e-03   0.180 0.856923    
lgb_pred:svm_pred          4.003e-03  1.813e-03   2.208 0.027376 *  
rf_pred:lgb_pred:svm_pred -3.060e-05  8.728e-06  -3.506 0.000468 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 51.6 on 1532 degrees of freedom
Multiple R-squared:  0.531, Adjusted R-squared:  0.5288 
F-statistic: 247.8 on 7 and 1532 DF,  p-value: < 2.2e-16

Coverages

• 0 coverages:

\(n\) and \(p\)

• 2204 observations
  • 1542 training
  • 662 testing
• 90 predictors
  • PC1, PC2, PC3, PC4, PC5, PC6, PC7, PC8, PC9, PC10, PC11, PC12, PC13, PC14, PC15, PC16, PC17, PC18, PC19, PC20, PC21, PC22, PC23, PC24, PC25, PC26, PC27, PC28, PC29, PC30, PC31, PC32, PC33, PC34, PC35, PC36, PC37, PC38, PC39, PC40, PC41, PC42, PC43, PC44, PC45, PC46, PC47, PC48, PC49, PC50, PC51, PC52, PC53, PC54, PC55, PC56, PC57, PC58, PC59, PC60, PC61, PC62, PC63, PC64, PC65, PC66, PC67, PC68, PC69, PC70, PC71, PC72, PC73, PC74, PC75, PC76, PC77, PC78, PC79, PC80, PC81, PC82, PC83, PC84, PC85, PC86, PC87, PC88, PC89, PC90

Component Models

• Tuning used 5-fold CV
• Final hyperparameters:

Random forest:

$num.trees
[1] 1000

$mtry
[1] 35

$min.node.size
[1] 2

$sample.fraction
[1] 0.9

$replace
[1] FALSE

$formula
agb_mgha ~ .

LGB:

$nrounds
[1] 500

$params
$params$learning_rate
[1] 0.05

$params$num_leaves
[1] 16

$params$max_depth
[1] 12

$params$extra_trees
[1] TRUE

$params$min_data_in_leaf
[1] 18

$params$bagging_fraction
[1] 0.3

$params$bagging_freq
[1] 0

$params$feature_fraction
[1] 0.5

$params$min_data_in_bin
[1] 2

$params$lambda_l1
[1] 4

$params$lambda_l2
[1] 0.3

$params$force_col_wise
[1] TRUE

SVM:

$x
agb_mgha ~ .

$kernel
[1] "laplacedot"

$type
[1] "eps-bsvr"

$kpar
$kpar$sigma
[1] 0.00390625


$C
[1] 44

$epsilon
[1] 0.0009765625

$nu
[1] 0.2