CassavAI: Cassava Leaf Disease Detection using Neural Networks

Dennis Dominic Diego, Jason Dolorso, Cymon Marcaida, Naman Punit, Matthew Romero
Asian Institute of Management

Abstract

In the Philippines, Cassava is considered one of the staple crops because of its ease in cultivation and profitability. However, the threat of various diseases hampers the growth and volume of cassava harvest which could be used for livelihood or exportation. It is imperative then that the diseases be detected at an early stage. This however falls into a subjective factor, as farmers or scientists often rely on optical inspection and domain knowledge whether a specific disease is one or the other. Since most cassava diseases manifest on leaves, we propose a deep learning method to detect cassava leaf diseases using neural networks, specifically pre-trained neural networks. Utilizing transfer learning on an EfficientNetB3 variant and training on 10,791,220 parameters, our model was able to achieve an accuracy of 88.60%. After applying data augmentations and ensemble methods, we were able to achieve an accuracy of 89.90%, beating the baseline proportional chance criterion of 50.46%. Despite the imbalanced data, we were able to create a neural network classifier that could accurately classify cassava leaf diseases which could potentially help farmers and agricultural scientists alike. In this project, we also outlined our recommendations to further the efforts in classifying plant leaf diseases and advance smart agriculture in the country.

Keywords

Cassava disease detection; Convolutional neural networks; Efficientnet; Transfer learning

---

Full text article and source codes can be provided upon request.