Classification Modeling Method for Near-Infrared Spectroscopy of Tobacco Based on Multimodal Convolution Neural Networks

Table 5

Model classification accuracy with different 2-D image input sizes.

Model	2-D NIRS image sizes	Feature spectrum range of 1-D NIRS (approximate)	Test set classification accuracy (%)
LeNet-5	40 ∗ 40	4000–10000 cm^−1	81.67
	38 ∗ 38	4000–9360 cm^−1	87.56
	36 ∗ 36	4000–8800 cm^−1	90.81
	34 ∗ 34	4000–8250 cm^−1	92.47
	32 ∗ 32	4000–7740 cm^−1	93.05
	30 ∗ 30	4000–7260 cm^−1	91.83
	28 ∗ 28	4000–6810 cm^−1	84.46
	26 ∗ 26	4000–6400 cm^−1	78.72