Performance of non-parametric algorithms for spatial mapping of tropical forest structure

Table 1 Machine learning performance using different input layers

Input layers	L only	A only	S only	L + A	L + A + S	L + A + S + T	BC
RF
RMSE	6.02 ± 0.10	5.80 ± 0.08	6.20 ± 0.04	5.41 ± 0.07	5.06 ± 0.07	4.51 ± 0.07	4.58 ± 0.09
R²	0.33 ± 0.02	0.39 ± 0.02	0.30 ± 0.01	0.46 ± 0.01	0.53 ± 0.01	0.63 ± 0.01	0.63 ± 0.01
MSD	0.08 ± 0.23	−0.13 ± 0.33	−0.16 ± 0.31	0.10 ± 0.22	0.12 ± 0.22	0.08 ± 0.25	−0.08 ± 0.18
MSD1	3.68 ± 1.20	2.32 ± 1.15	6.84 ± 1.25	3.80 ± 1.21	3.89 ± 1.24	4.48 ± 0.99	0.71 ± 0.81
MSD2	−8.50 ± 0.45	−7.63 ± 0.60	−7.03 ± 0.81	−7.46 ± 0.50	−5.61 ± 0.59	−5.16 ± 0.43	−1.73 ± 0.57
ME
RMSE	6.00 ± 0.10	5.62 ± 0.06	6.12 ± 0.05	5.46 ± 0.09	5.17 ± 0.09	4.73 ± 0.13	5.29 ± 0.17
R²	0.34 ± 0.02	0.42 ± 0.01	0.32 ± 0.01	0.46 ± 0.02	0.52 ± 0.01	0.59 ± 0.02	0.53 ± 0.02
MSD	0.10 ± 0.23	−0.12 ± 0.31	−0.16 ± 0.34	0.10 ± 0.24	0.05 ± 0.25	0.02 ± 0.24	−0.15 ± 0.17
MSD1	5.76 ± 0.71	4.27 ± 1.04	9.00 ± 1.33	4.71 ± 0.67	4.71 ± 0.56	4.57 ± 0.38	1.10 ± 0.64
MSD2	−10.88 ± 0.50	−9.88 ± 0.69	−8.24 ± 1.18	−8.78 ± 0.90	−6.73 ± 0.73	−4.39 ± 0.24	−1.13 ± 0.34

The sample size for this test was fixed at 400 samples, and the rest 32,674 100-m pixels were used as test samples. The results were cross validated by repeated random sampling of the training data (Monte Carlo CV). RF and ME predictions were evaluated using RMSE, R², overall MSD, MSD for small trees (MSD1) and the MSD for large trees (MSD2). The input “L only” includes four Landsat bands, “A only” uses only the two ALOS bands, “S only” uses only the SRTM bands, “L + A” includes four Landsat and two ALOS bands, “L + A + S” includes Landsat, ALOS and SRTM bands, “L + A + S + T” includes all satellite bands plus texture layers, and “BC” uses the same set of input layers as “L + A + S + T”, but results are from the bias-corrected algorithms

ISSN: 1750-0680