Early Diagnosis of Soybean Disease using Pretrained Network: A Comparative Study of EfficientNetV2 and MobileNetV2 with Advanced Feature Extraction Technique

Diseases affecting the soybean crop are widespread and have an impact on both the amount and quality of yield. Farmers will profit from early illness identification utilising a quick, dependable, nondestructive approach to combat the issue. In this work, two deep learning-based architectures, EfficientNetV2 and MobileNetV2, are fed pictures of soybean leaves—six diseased and one healthy class—obtained from the NCSRP (North Central Soybean Research Programme) dataset. Analysis has been done on how the quantity of pictures and the importance of the hyperparameters—minibatch size, weight, and bias learning rate—affect the execution time and accuracy of the classification. Additional research used Mobilenetv2 and EfficientNetV2 to examine the efficacy of different feature extraction methods for the categorization of soybean diseases. K stands for clustering for segmentation it also refers to a combination of techniques that combines colour, texture, and form features, contour detection, Canny edge detection; colour thresholding for background removal, and Grey Level Co-occurrence Matrix (GLCM) for grayscale analysis. The last layer was substituted with an output layer, a softmax layer, and a fully linked layer in order to categorise seven classes—six sick and one healthy. Based on preliminary classification findings, EfficientNetV2 achieves 97.49% accuracy, while MobileNetV2 achieves 97.23% accuracy. 17,938 photos were used to get these findings. Additionally, MobileNetV2 yields 96.19% accuracy and Efficient NetV2 yields 95.81% accuracy after assigning the same number of photos to each class, 373. Minibatch size research reveals that MobileNetV2 achieves 96.51% with ( minibatch size = 12), while EfficientNetV2 achieves the greatest accuracy of 99.24% with a (minibatch size= 2).Greater minibatch sizes result in faster EfficientNetV2 and slower MobileNetV2 execution times. When the batch size exceeded 32, EfficientNetV2 failed because of out-of-memory problems. Additional learning rate investigation reveals that when learning rate is increased, EfficientNetV2's accuracy decreases from 97.33% to 96.19%. MobileNetV2 accuracy declines until a learning rate of 30, it reaches 96.38% at a rate of 40. On keeping minibatch size =2, and a learning rate ten times lower than the global learning rate, EfficientNetV2 outperformed. The feature that was retrieved and utilized to train and assess two models' performance. The outcome shows that EfficientNetV2 consistently performs better than MobileNetV2 when it comes to various feature extraction techniques, with segmented features yielding the best accuracy. According to the data, EfficentNetV2 constantly performs better than MobileNetV2 when it comes to various feature extraction techniques. With segmented features, it achieved the greatest accuracy of 99.24%, while MobileNetV2 only managed 96.51%.