Deep learning based high-throughput phenotyping of chalkiness in rice exposed to high night temperature

Table 3 Classification results on polished rice with various networks as backbone in the weakly supervised Grad-CAM approach

Model	Acc.(%)	Chalky			Non-chalky
Model	Acc.(%)	Pre.(%)	Rec.(%)	F1(%)	Pre.(%)	Rec.(%)	F1(%)
DenseNet-121	95.61	94.58	94.58	94.58	96.31	96.31	96.31
DenseNet-161	95.12	92.44	95.78	94.08	97.06	94.67	95.85
DenseNet-169	94.63	92.86	93.98	93.41	95.87	95.08	95.47
ResNet-18	94.63	94.44	92.17	93.29	94.76	96.31	95.53
ResNet-34	94.15	93.29	92.17	92.73	94.72	95.49	95.10
ResNet-50	94.88	95.03	92.17	93.58	94.78	96.72	95.74
ResNet-101	95.12	93.45	94.58	94.01	96.28	95.49	95.88
ResNet-152	94.88	93.94	93.37	93.66	95.51	95.90	95.71
SqueezeNet-1.0	95.12	93.45	94.58	94.01	96.28	95.49	95.88
SqueezeNet-1.1	94.39	91.33	95.18	93.22	96.62	93.85	95.22
VGG-11	94.88	93.94	93.37	93.66	95.51	95.90	95.71
VGG-13	94.39	92.31	93.98	93.13	95.85	94.67	95.26
VGG-16	95.12	92.94	95.18	94.05	96.67	95.08	95.87
VGG-19	94.15	90.34	95.78	92.98	97.01	93.03	94.98
EfficientNetB0	95.13	93.98	93.98	93.98	95.92	95.92	95.92
EfficientNetB1	95.13	94.51	93.37	93.94	95.55	96.33	95.93
EfficientNetB2	93.67	90.23	94.58	92.35	96.20	93.06	94.61
EfficientNetB3	95.13	95.06	92.77	93.90	95.18	96.73	95.95
EfficientNetB4	95.38	96.82	91.57	94.12	94.49	97.96	96.19
EfficientNetB5	93.67	91.67	92.77	92.22	95.06	94.29	94.67
EfficientNetB6	94.16	92.77	92.77	92.77	95.10	95.10	95.10

The number following a network’s name denotes the number of layers in the network (as in DenseNet-121 or ResNet-101) or the version of the network (as in SqueezeNet-1.0 or EfficientNetB0). Performance is reported in terms of Accuracy (Acc.), Precision (Pre.), Recall (Rec.) and F1 measure (F1). Precision, Recall and F1 measure values are reported separately for the Chalky and Non-Chalky classes. All models are trained/tuned/evaluated on the same training/development/test splits. The results reported are obtained on the test set. The best performance for each type of model for each metric is highlighted using bold font

ISSN: 1746-4811