Impact of Data Centric Approach to Improve the Performance of Leaf Disease Classification

Abstract

Cassava is the third highest carbohydrate food after rice and maize. Due to various plant diseases found in Cassava, security threat is posed in developing countries. To stop spoiling the whole plant, a lot of effort has been made by researchers to identify early leaf disease as agricultural farming is closely associated with every country’s economy. Recently, Machine Learning (ML) models have achieved a lot of success with big data, available resources and improvement in learning algorithms. However, there are invalid instances in big data like ambiguous or mislabelled or irrelevant. Though the performance of ML model is not degraded if such instances are small as the effect of the average gradient is small during the training. But the performance of system is degraded if quantity of invalid instances is large. The existing work for leaf disease detection is performed using Model Centric approach, where hyperparameter tuning is performed to enhance the performance of the system. Recently, focus in changed from Model Centric to Data Centric approach, where model is fixed but quality of dataset is enhanced by considering the consistency of labels, systematic sampling of training data and selection of appropriate batches. This is an invaluable step towards improvement of any system. In this work, noise label detection and correction are performed on Cassava Leaf Disease Classification dataset using confidence learning. The generated quality data is given to the model, and performance comparison is made between model centric and data centric approaches, which concludes that the performance of data centric is improved by 6.33%. The conclusion of this work is not to downgrade the significance of model centric, but to showcase the neglected potential of enhancing the performance of such systems using data centric approach.