Data augmentation for automated pest classification in Mango farms

Mangos are native to South and Southeast Asian regions. They are one of the favorite fruits consumed globally, with an overall estimated consumption reaching up to 50.65 million metric tons in 2017. However, the production of mango is usually severely affected by pests that attack the fruit, stem, root or mango leaf. Addressing the need for an early stage automated or semi-automated pest identification system, the research presented in this paper proposes an advanced machine learning (ML) technique for analyzing large-scale mango fields and identification of the onset of biological threats using computer vision and deep-learning technologies. The ML technique presented in the paper extends the pre-trained VGG-16 deep-learning model to supplement the last layer with a fully connected network training of consisting of 2-layers. In addition, the research presented in the paper also considers the real-world operational conditions commonly faced by Indonesian farmers for collecting and processing visual information obtained from the Mango farms. The sparsity of the dataset availability for effectively training deep-learning network is addressed through the application of data augmentation process that is able to accurately recreate the conditions faced by the farmers. The overall accuracy of the proposed training solution achieved is 73% on the validation dataset and 76% for the testing data. The application of the augmentation transformation function leads to an improvement of 13.43% of accuracy on the testing data.