A comparative analysis of basic convolutional neural network against MobileNetV3 model for tomato leaf disease classification on primary field data in Malawi

Abstract

Malawi is an Agro based economy reliant on agriculture for both commercial and subsistence farming. The production of tomatoes in Malawi, led by smallholder and small medium enterprises (SMEs), is insufficient to satisfy local demand due to issues such pests, illnesses, unstable markets, and high input costs. A majority of mainstream farmers who undertake the practice do not have the adequate technical knowledge to recognize and efficiently control diseases and pests. To address the problem of tomato leaf disease identification, this research aimed to develop an automated system for tomato leaf disease detection by utilizing data augmentation techniques, MobileNetV3 and basic CNN algorithms. Models were trained on secondary data collected from the public PlantVillage dataset, and the resultant classifiers were tested on primary data from local farm images. The research was able to determine that through the results discovered, proposed data augmentation methods based on geometric and colour transformations may improve the performance of models but not significantly especially in the case of models that detect tomato leaf diseases. The experimental results demonstrate that both models -basic CNN and pretrained MobileNetV3- that were trained on the augmented dataset performed better than those trained on the non-augmented dataset. Additionally, with an accuracy of 92.59% and a loss of 0.2805, the pre-trained MobileNetV3 model that was trained on augmented data conventionally performs better than a basic CNN model. However, when tested on the primary field dataset, the models did not meet expectations for generalization, with the pretrained MobileNetV3 achieving an accuracy of 9.2%, and loss of 12.91 and the basic CNN achieving an accuracy of 10.14%, and loss of 8.11. The experiments aided in showing that the models that are trained on PlantVillage dataset are not as effective when they are used in real world scenarios. Further improvements are needed to enhance the models' generalization in real-world scenarios.