Enhancing Machine Learning Training Performance in Smart Agriculture Datasets Using a Mobile App

The agricultural sector faces significant challenges due to invasion of pests that damage crops and cause significant loss of production. Traditional methods to detect these insects are cost ineffective, thus, automated vision systems based on machine learning (ML) have recently been proposed in the literature. However, a significant issue is the lack of a prior dataset to build the ML model on. To mitigate this problem, we propose a new approach to train a model using a small initial dataset and continually improve the accuracy process by retraining it on new images labeled by a mobile application. Retraining is performed on new data, which comes from a mobile application that displays pictures of insects and prompts expert users to label them. The users' input is used to retrain the model on new coming images. Specifically, our method trains the model on 100 initial images, and retrains it with every 100 new images. The IP102 large-scale dataset for pest recognition was used to demonstrate the effectiveness of the approach. The results show an improvement of accuracy of up to 50 percentage units for the built Convolutional Neural Network (CNN) model.