Enhancing cassava disease detection using CNN models trained from scratch: A comparative study with transfer learning approaches

Abstract

Agriculture is a vital sector, with farmers playing a crucial role in producing food and supporting the global population. However, plant diseases pose significant challenges, leading to substantial crop yield losses and threatening food security., Artificial intelligence models, particularly convolutional neural networks (CNNs), are widely used for object identification and image classification. Although transfer learning techniques help reduce computation time, CNNs often struggle to differentiate between highly similar images because of the limited flexibility in configuring and fine-tuning parameters. To address issues related to random configuration, feature selection, and model architecture, we present a CNN model with 15 layers, trained from scratch. This approach allows for a setup and design better suited to the task of distinguishing similar images. The proposed model is compared to two other CNN-based transfer learning models, InceptionV3 and VGG16, which are trained solely on the top layers with fixed bottom layers. Experimental results indicate that the proposed model outperformed the other models by 24%, achieving an accuracy of 84% over 50 epochs. The experimental results suggest that the proposed approach can achieve even better performance with additional training epochs, as the model's training graph was still improving.