Deep Learning Application for Predicting Soil Organic Matter Content by VIS-NIR Spectroscopy
Table 1
Prediction accuracy results of different model with the fractional-order derivative (1.5) transformation of soil spectral reflectance data.
Model
Variables
Training
Validation
R2
RMSE
R2
RMSE
RPD
BPN
67
0.831 ± 0.003
4.900 ± 0.062
0.867 ± 0.000
5.382 ± 0.001
2.843 ± 0.001
BPN
203
0.835 ± 0.018
5.107 ± 0.290
0.834 ± 0.018
6.112 ± 0.290
2.465 ± 0.092
MLP
67
0.806 ± 0.021
5.657 ± 0.288
0.861 ± 0.019
5.578 ± 0.364
2.709 ± 0.164
MLP
203
0.890 ± 0.011
4.249 ± 0.199
0.886 ± 0.002
5.044 ± 0.037
2.987 ± 0.022
LeNet5
67
0.872 ± 0.015
4.542 ± 0.224
0.870 ± 0.020
5.304 ± 0.401
2.901 ± 0.220
LeNet5
203
0.902 ± 0.006
3.909 ± 0.114
0.863 ± 0.001
5.471 ± 0.033
2.790 ± 0.022
DenseNet10
67
0.927 ± 0.009
3.376 ± 0.191
0.888 ± 0.012
4.925 ± 0.274
3.133 ± .0.177
DenseNet10
203
0.907 ± 0.009
3.367 ± 0.177
0.892 ± 0.004
4.933 ± 0.091
3.053 ± 0.056
Data are the mean ± standard deviation. R2 = coefficient of determination, RMSE = root mean square error, and RPD = the ratio of performance to deviation.