Deep Learning Application for Predicting Soil Organic Matter Content by VIS-NIR Spectroscopy
Table 2
Prediction accuracy results of using different depths of DenseNet based on the full-spectrum data without the fractional-order derivative transformation.
Model
Learning rate
Training
Validation
R2
RMSE
R2
RMSE
RPD
DenseNet10
0.001
0.832 ± 0.012
5.334 ± 0.177
0.806 ± 0.007
6.519 ± 0.128
2.345 ± 0.050
DenseNet10
0.0006
0.837 ± 0.013
5.188 ± 0.203
0.836 ± 0.010
6.037 ± 0.199
2.508 ± 0.085
DenseNet13
0.001
0.878 ± 0.010
4.376 ± 0.164
0.833 ± 0.005
6.117 ± 0.095
2.462 ± 0.038
DenseNet13
0.0006
0.861 ± 0.012
4.697 ± 0.199
0.828 ± 0.006
6.207 ± 0.108
2.431 ± 0.044
DenseNet16
0.001
0.856 ± 0.012
4.782 ± 0.197
0.827 ± 0.008
6.172 ± 0.135
2.461 ± 0.051
DenseNet16
0.0006
0.858 ± 0.011
4.792 ± 0.177
0.848 ± 0.007
5.847 ± 0.131
2.576 ± 0.058
DenseNet19
0.001
0.870 ± 0.010
4.515 ± 0.179
0.853 ± 0.007
5.722 ± 0.124
2.639 ± 0.056
DenseNet19
0.0006
0.916 ± 0.009
3.628 ± 0.177
0.853 ± 0.011
5.745 ± 0.222
2.622 ± 0.099
Data are the mean ± standard deviation. R2 = coefficient of determination, RMSE = root mean square error, and RPD = the ratio of performance to deviation.
