A New Local Modelling Approach Based on Predicted Errors for Near-Infrared Spectral Analysis

Table 3

Performance comparisons among local errors regression, global method, other local methods with different similarity criterions.

Method	Similarity criterion	RMSEP		RPD	Size of subset	Parameters	Time consumption (s)
Global	—	4.30	0.96	5.11	460	—	4.4
Other local methods	ED	4.18	0.96	5.26	150	—	175.4
	Cosine	4.25	0.96	5.17	150	—	179.1
	PC-M	4.21	0.96	5.22	50	PC factors = 10	53.5
	+ + ED	4.24	0.96	5.18	100	γ = 0.8	123.1
	+ + SLPP	4.27	0.96	5.15	200	γ = 0.8, = 20	233.9
Local errors regression	Errors + ED	3.21	0.98	6.85	13~205	= 20	9.5

ED: Euclidean distance; PC-M: Principal components-Mahalanobis distance; + + ED: Euclidean distance considering both spectra and property ; + + SLPP: Euclidean distance in the low-dimensional space obtained with supervised locality preserving projection method; errors + ED: Euclidean distance between predicted errors; RMSEP: root mean squared error of prediction; : correlation coefficient in prediction set; RPD: residual prediction deviation; PC factors: Principal component factors; symbol : a trade-off parameter to balance the importance of spectra and property ; : dimension of transformation matrix; and s: second.