Landsat:LiDAR AGB 1.1.0 More data = more better

Change Summary

Last iteration: 2022-04-11 - Landsat AGB 0.0.99: Final Countdown

Change Summary

• This version adds 65 indices and their respective 1 year deltas (130 predictors) computed through awesome spectral indices library.

• Document includes comparisons against Landsat-FIA test plots

LiDAR-AGB Test Pixel Results

	RF (ranger)	GBM (LightGBM)	Ensemble (model weighted)	Ensemble (RMSE weighted)
%RMSE	25.989	24.215	23.963	24.179
RMSE	36.597	34.099	33.745	34.049
MAE	28.644	26.288	25.846	26.428
MBE	-0.041	0.043	-0.049	0.002
R2	0.801	0.824	0.827	0.825

Landsat-FIA Test Plot Results

	RF (ranger)	GBM (LightGBM)	Ensemble (model weighted)	Ensemble (RMSE weighted)
%RMSE	41.001	40.263	39.928	39.778
RMSE	56.324	55.311	54.850	54.645
MAE	44.004	41.756	41.554	41.602
MBE	5.598	-9.391	-5.370	-2.145
R2	0.326	0.362	0.361	0.358

Landsat-FIA (With LiDAR-identified Zeroes) Test Plot Results

	RF (ranger)	GBM (LightGBM)	Ensemble (model weighted)	Ensemble (RMSE weighted)
%RMSE	43.648	41.794	41.944	42.165
RMSE	55.161	52.818	53.007	53.286
MAE	42.976	40.570	40.660	41.212
MBE	7.721	3.034	3.769	5.300
R2	0.473	0.508	0.506	0.503

LiDAR-AGB Pixel Distribution

summary(bind_rows(training, testing)$agb_mgha)

     Min.   1st Qu.    Median      Mean   3rd Qu.      Max. 
  0.00721  70.68408 141.33738 140.87515 212.00599 279.53638 

Metadata

Ensembles

• RMSE-weighted model weights:

• Linear model weights:

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

Residuals:
     Min       1Q   Median       3Q      Max 
-173.449  -20.012   -0.395   20.508  154.379 

Coefficients:
                   Estimate Std. Error t value Pr(>|t|)    
(Intercept)      -3.332e+00  1.306e+00  -2.551   0.0108 *  
rf_pred           3.173e-01  2.263e-02  14.019   <2e-16 ***
lgb_pred          6.788e-01  2.110e-02  32.164   <2e-16 ***
rf_pred:lgb_pred  1.548e-04  7.071e-05   2.189   0.0286 *  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 33.87 on 9996 degrees of freedom
Multiple R-squared:  0.8257,    Adjusted R-squared:  0.8257 
F-statistic: 1.579e+04 on 3 and 9996 DF,  p-value: < 2.2e-16

\(n\) and \(p\)

• observations
• 10000 training
• testing
• 166 predictors
  • precip, tmin, tmax, dem, twi, slope, aspect, chm, aoa, yod, mag, normg, normnir, normr, nsds, nwi, nrfig, nli, ngrdi, nirv, ndvi, ndwi, ndvimndwi, nddi, ndsi, mtvi2, muwir, nbr2, nbsims, mndwi, mrbvi, msavi, msi, msr, s3, si, sr, sr2, swi, rcc, pisi, rdvi, rgbvi, tgi, tvi, trivi, rgri, tdvi, vgnirbi, ui, vig, aweinsh, aweish, bai, bcc, dvi, vrnirbi, wi1, wi2, embi, gli, gemi, gndvi, grndvi, gosavi, exr, exgr, gbndvi, gari, fcvi, gvmi, mbi, lswi, ikaw, ipvi, nmdi, mcari2, mlswi26, mlswi27, nbr, tcg, tcb, tcw, delta_normg, delta_normnir, delta_normr, delta_nsds, delta_nwi, delta_nrfig, delta_nli, delta_ngrdi, delta_nirv, delta_ndvi, delta_ndwi, delta_ndvimndwi, delta_nddi, delta_ndsi, delta_mtvi2, delta_muwir, delta_nbr2, delta_nbsims, delta_mndwi, delta_mrbvi, delta_msavi, delta_msi, delta_msr, delta_s3, delta_si, delta_sr, delta_sr2, delta_swi, delta_rcc, delta_pisi, delta_rdvi, delta_rgbvi, delta_tgi, delta_tvi, delta_trivi, delta_rgri, delta_tdvi, delta_vgnirbi, delta_ui, delta_vig, delta_aweinsh, delta_aweish, delta_bai, delta_bcc, delta_dvi, delta_vrnirbi, delta_wi1, delta_wi2, delta_embi, delta_gli, delta_gemi, delta_gndvi, delta_grndvi, delta_gosavi, delta_exr, delta_exgr, delta_gbndvi, delta_gari, delta_fcvi, delta_gvmi, delta_mbi, delta_lswi, delta_ikaw, delta_ipvi, delta_nmdi, delta_mcari2, delta_mlswi26, delta_mlswi27, delta_nbr, delta_tcg, delta_tcb, delta_tcw, lcpri3, lcpri4, lcpri6, lcpri2, lcsec1, lcsec3, lcsec4, lcsec2, lcsec8, lcsec5, lcsec6

Component Models

• Tuning used 5-fold CV
• Final hyperparameters:

Random forest:

$num.trees
[1] 2750

$mtry
[1] 54

$min.node.size
[1] 4

$sample.fraction
[1] 1

$replace
[1] FALSE

$formula
agb_mgha ~ .

LGB:

$params
$params$num_leaves
[1] 40

$params$max_depth
[1] 8

$params$extra_trees
[1] TRUE

$params$min_data_in_leaf
[1] 10

$params$bagging_fraction
[1] 0.3

$params$bagging_freq
[1] 0

$params$feature_fraction
[1] 0.6

$params$min_data_in_bin
[1] 13

$params$lambda_l1
[1] 7

$params$lambda_l2
[1] 8

$params$learning_rate
[1] 0.1

$params$force_col_wise
[1] TRUE

$params$nrounds
[1] 1000