Er demonstrates the outstanding performance of CNNs in maize leaf disease detection by comparing the accuracy of lots of CNNs, such as AlexNet, VGG19, ResNet50, DenseNet161, GoogLeNet, and their optimized versions primarily based on MAF module, with conventional machine understanding algorithms, SVM [24] and RF [25]. The comparison benefits are shown in Table 3.Table three. Accuracy of diverse models. Model SVM RF baseline MAF-AlexNet baseline MAF-VGG19 baseline MAF-ResNet50 baseline Chloramphenicol palmitate Bacterial MAF-DenseNet161 baseline MAF-GoogLeNet Tanh ReLU LeakyReLU Sigmoid Mish Accuracy 83.18 87.13 92.82 93.11 93.49 92.80 93.92 94.93 95.30 95.18 95.08 95.93 97.41 96.18 96.18 95.90 96.75 97.01 94.27 95.01 95.09 94.27Remote Sens. 2021, 13,15 ofThe benefits of experiments indicate that the accuracy of your mainstream CNNs might be enhanced with the MAF module, and the impact on the ResNet50 stands out, reaching two.33 . Also, it is also located that the promoting impact of adding all activation functions towards the MAF module is not the best. As an alternative, the combination of Sigmoid, ReLU (or tanh), and Mish (or LeakReLU) ranks leading. 3.2.1. Ablation Experiments to Confirm the Effectiveness of Warm-Up Ablation experiments have been performed on various models to confirm the validation of the warm-up approach. The outcomes are shown in Figure 17.Figure 17. Loss curve of unique models and strategies.three.two.two. Ablation Experiments To confirm the effectiveness of the many pre-processing strategies proposed within this write-up, like distinctive information augmentation techniques, the ablation experiments have been performed on MAF-ResNet50, chosen from the above experiments with the finest performance. The experimental final results are shown in Tables four and 5.Table 4. Ablation experiment result of different pre-processing strategies.Removal of Facts baselineGray-ScaleSnapmixMosaicAccuracy 95.08 97.41 96.29 95.82 93.17 94.39MAF-ResNetTable 5. Ablation experiment outcome of other techniques. DCGAN baseline MAF-ResNet50 D-Isoleucine Protocol LabelSmoothing Bi-Tempered Loss Accuracy 95.08 96.53 97.41 95.77 97.22Remote Sens. 2021, 13,16 ofThrough the evaluation of experimental results, we are able to discover these data enhancement methods like Snapmix and Mosaic are of excellent assistance in enhancing the functionality of the MAF-ResNet50 model. The principles of Snapmix and Mosaic are comparable. It might be noticed that the model performs most effective when warm-up, label-smoothing, and Bi-Tempered logistic loss solutions are employed simultaneously, as shown in Table 5. 4. Discussion 4.1. Visualization of function Maps Within this paper, the output of multi-channel function graphs corresponding to eight convolutional layers of the MAF-ResNet50 was visualized with the highest accuracy in the experiment, as shown in Figure 18. As can be seen in the figure, inside the shallow layer feature map, MAF-ResNet50 extracted the lesion details of your maize stalk lesion and carried out depth extraction inside the subsequent feature map. As the network layer deepened, the interpretability of the function map visualization became worse. Nevertheless, even in Figure 19, the corresponding connection among the highlighted colour block area of your feature map and the lesion location in the original image can nonetheless be observed, which further reveals the effectiveness of the MAF-ResNet50 model.Figure 18. Visualization of shallow function maps.Figure 19. Visualization of your deep feature map.Remote Sens. 2021, 13,17 of4.two. Intelligent Detection Technique for Maize Diseases To confirm the robus.