Table 12 Selected hyperparameters of the convolution neural network models using the signature bands with the lowest prediction error in each spectral range.

From: Hyperspectral signature-band extraction and learning: an example of sugar content prediction of Syzygium samarangense

Spectral range (nm)

400–1700

400–1000

400–700

Bandwidth (nm)

± 10

± 7.5

± 12.5

Conv. kernel size

4 × 1

4 × 1

4 × 1

Avg. pooling size

2 × 2

2 × 2

1 × 4

FC1 units

1024

1024

1024

FC2 units

512

512

512

FC3 units

1

1

1

Batch size

16

128

64

Learning rate

0.0004

0.0008

0.0004

Back to article page