Improving the performance of a spectral model to estimate total nitrogen content with small soil samples sizes

Chemical and Biological Technologies in Agriculture

Table 2 Estimation results of using different fusion strategies and sub-regressors

Fusion strategy	Regressor	Calibration		Validation		Tenfold validation
		\(R^2_c\)	\(RMSE_c(g.kg^{-1})\)	\(R^2_v\)	\(RMSE_v(g.kg^{-1})\)	\(R^2_{cv}\)	RPIQ
Single^[1]	GPR^[2]	0.695	0.140	0.627	0.107	0.598	2.895
Bagging	GPR	0.702	0.138	0.711	0.088	0.651	2.942
Stacking	GPR	0.711	0.131	0.723	0.085	0.653	2.951
Single	UVE-PLSR^[3]	0.645	0.156	0.597	0.113	0.550	2.937
Bagging	UVE-PLSR	0.715	0.140	0.737	0.083	0.693	3.324
Stacking	UVE-PLSR	0.734	0.123	0.714	0.087	0.677	3.340
Single	SPA-MLR^[4]	0.429	0.198	0.553	0.117	0.484	2.501
Bagging	SPA-MLR	0.683	0.147	0.709	0.087	0.680	3.314
Stacking	SPA-MLR	0.773	0.125	0.784	0.075	0.720	3.511

\([1]\) Single implies without using any model fusion strategy, a single model is directly implemented on the expanded dataset
\([2]\) GPR represents using Gaussian progress regressor to build sub-models
\([3]\) UVE-PLSR is to use the combination of uninformative variable elimination techniques and partial least regressor to build sub-models
\([4]\) SPA-MLR is to implement the successive projections algorithm to remove co-linear variables and establish multiple linear regression sub-models