Evaluating CNN and Vvsion transformer models for Mango Leaf variety identification

Abstract

Mango, often referred to as the “King of fruits”, occupies a superior place in the global agricultural landscape due to its growing demand. Thus accurate identification and classification of mango tree varieties is essential to improve quality control and inventory management in this context. In this study, we harness the power of well-established deep learning models, to detect the type and variety of mango leaves by using the mango leaf image processing method. Our meticulous analysis of accuracy and loss curves provides insight into model performance, ensuring the model is not overfitted. Additionally, we construct a comprehensive confusion matrix, highlighting the system’s ability to distinguish between different mango tree varieties. We also introduced a detailed classification report, offering precision, recall, F1 score, and support for each mango tree variety. This report is a valuable tool for stakeholders, helping them make informed decisions about quality control and inventory management. Notably, we curated a vast dataset of 14,000 raw mango leaf images, collected from different locations and seasons, reflecting the diversity of mango cultivation. Our database contains 26 types of different mango leaf variants. In the proposed system, various Deep Learning and Machine Learning algorithms were utilized including VGG16, EfficientNetB3, MobileNetV2, InceptionV3, Xception, ResNet50 and ViT for classification, and a comparison was made based on their accuracy rate which is respectively 98.64%, 87.19%, 97.90%, 98.89%, 98.42%, 98.10% & 97%. By combining precision curves, loss curves, confusion matrices, and classification reports, we provide a comprehensive performance evaluation of our system. This work will bring a cathartic change in our agricultural economy by easing the process of identifying mango plants.