Using Deep Convolutional Neural Networks for Image-Based Diagnosis of Nutrient Deficiencies in Rice

Table 5

Convolutional neural networks and specifications for experiments. Accuracy and kappa score of training, validation, and test are also included.

Model	Parameters (millions)	Layers	Training accuracy (%)	Validation accuracy (%)	Test accuracy (%)	Validation kappa score	Test kappa score	Training time per epoch (1090 samples)	Validation time per epoch (364 samples)
Inception-v3	23.9	48	98.75 ± 0.19	93.86 ± 0.57	91.67 ± 0.63	0.9314 ± 0.0064	0.9070 ± 0.0070	34 s	2 s
ResNet50	26.7	50	99.81 ± 0.14	97.53 ± 0.01	95.15 ± 0.16	0.9723 ± 0.0002	0.9478 ± 0.0022	30 s	2 s
NasNet-Large	84.9	—	99.79 ± 0.19	95.88 ± 0.95	96.25 ± 0.57	0.9539 ± 0.0106	0.9611 ± 0.0093	199 s	10 s
DenseNet121	8.1	121	99.30 ± 0.05	98.62 ± 0.57	97.44 ± 0.57	0.9794 ± 0.0064	0.9713 ± 0.0063	48 s	3 s
Color feature + SVM (RBF kernel)	—	—	90.55	66.48	64.01	0.6220	0.5951	—	—
HOG + SVM (RBF kernel)	—	—	93.76	56.93	56.86	0.5105	0.5064	—	—