Table 1. Identification accuracy training on BBFID-1 (scale A) at the genus level with different model architectures and hyperparameters.

Architectures in this table are shown in Fig. 3. “Trainable layers of functional layers” represents the size of the parameters that can be trained. “None” means that all layers of the backbone are frozen and the parameters involved in these layers cannot be trained. These parameters maintain the values at the time of model initialization. “Half layers” means that half of the backbone layer parameters are frozen, while “All layers” means that all parameters of this model are not frozen and can be updated during the training process. This setting has an impact on both the model training process and the model performance.

Order	Backbone	Batch size	Trainable layers of functional layers	Reduce LR on plateau	Epochs	Max. training accuracy	Min. training loss	Max. validation accuracy	Min. validation loss	Test accuracy	Test loss
1	VGG-16	32	None	Yes	50	0.8648	0.4212	0.6444	1.1440	0.6281	1.2512
2	VGG-16	32	Half layers	Yes	40	0.9959	0.0181	0.7515	0.9126	0.7330	0.8444
3	VGG-16	32	All layers	Yes	50	0.7670	0.6080	0.5698	1.3465	0.5386	1.4802
4	VGG-16	32	All layers	No	36	0.3609	1.8002	0.3338	2.0523	0.0957	3.0871
5	Inception-ResNet-v2	8	None	Yes	50	0.3236	1.9945	0.3385	2.0345	0.3225	2.1000
6	Inception-ResNet-v2	8	Half layers	Yes	50	0.7363	0.7163	0.5263	1.4931	0.4877	1.5584
7	Inception-ResNet-v2	8	All layers	Yes	46	0.9959	0.0216	0.7934	1.2041	0.7778	2.5044
8	Inception-ResNet-v2	8	All layers	No	46	0.9805	0.0602	0.7981	0.8178	0.6590	1.2590
9	EfficientNetV2s	8	None	Yes	50	0.5693	1.2799	0.5419	1.4210	0.4923	1.5424
10	EfficientNetV2s	8	Half layers	Yes	50	0.9708	0.1013	0.7624	0.8314	0.7515	0.8633
11	EfficientNetV2s	8	All layers	Yes	44	0.9959	0.0139	0.8338	0.6130	0.8302	0.6807
12	EfficientNetV2s	8	All layers	No	37	0.9825	0.0578	0.8136	0.7905	0.7886	0.8122