An Improved Model for Nutrient Deficiency Diagnosis of Rice Plant by Ensemble Learning

Abstract

Rice is the most popular staple food all over the world, especially in China, India, and Bangladesh. It is the primary food of Bangladesh. But its production is hampered by the nutrient imbalance in soil that results in failing to meet the required amount of rich needed for this country. Early detection of nutrient deficiency can improve rice production. Machine learning (ML) is one of the best solutions for early detection of nutrient shortfall by leaf image processing since the leaves change their physical appearance due to the nutrient scarcity of soil. In our work, we have focused on how ML performs better on the small dataset of Kaggle with 1156 images. Three pre-trained CNN models, MobileNet, DenseNet121, and DenseNet169, with an added pooling layer and dropout layer at the bottom, were carried out on an augmented dataset. The average ensemble of them outperforms and enhances testing accuracy from 92% to 96.67%. It yields a roc_auc score of 99.62%. The analysis in this work revealed that data augmentation with parameter tuning, transfer learning architecture, and ensemble learning play the key roles in improving accuracy.