Application of the Transfer Learning Method in Detecting Diseases in Strawberry Plants using the ESP32 Platform

Abstract

In this study, various models of deep convolutional artificial neural networks (CNNs) were explored to address the critical need for timely identification and prevention of plant diseases. Traditional CNN architectures generally involve high computational costs because they have a large number of parameters. To address this, the modified approach replaces standard convolutions with separable convolution, lowering the number of parameters and computational requirements. The models were trained on diverse datasets that included several plant species and disease classes. Evaluation of the model with different parameters such as batch size and dropout achieved impressive levels of disease classification accuracy: InceptionV3, InceptionResNetV2, MobileNetV2, and Effi- cientNetB0 recorded consecutive accuracy rates of 98.42%, 99.11%, 97.02%, and 99.56%, surpassing the handmade feature-based methods. In addition, G-ResNet50, a proposed model derived from ResNet50, enriched with focal loss, was introduced specifically for the identification of diseases in strawberries. Trained with an extended dataset through various operations, G-ResNet50 exhibits faster convergence and a much higher accuracy rate (98.67%) compared to VGG16, ResNet50, In- ceptionV3, and MobileNetV2. The G-ResNet50 model demonstrates high robustness, stability and recognition accuracy, presenting a practical solution for real-time detection and classification of strawberry diseases, which is essential for agricultural efficiency and productivity.