Abstract:
Bananas play a significant role in agriculture, especially in Sri Lanka, where
approximately 29 varieties are cultivated. Proper identification of the types
of bananas is essential for maintaining the best crop management practices,
including fertilization, irrigation, and control of disease aspects.
Identification of banana type by non-experts, done manually, may result in
inefficiencies and errors. This research is intended to automate the
identification of major banana species, such as Red Banana, Sour Banana,
Cavendish, and Ash Plantain, with the help of machine learning coupled with
image analysis. For this research, images were collected from multiple
banana plantations across different regions, focusing on key identifying
features such as petiole nature, color of the petiole, branching style, color of
the midrib, and leaf blade base. Including the above features, two image data
sets as branching style and leaf upper side per tree were collected. This data
has been split into 80% for training and 20% for testing to ensure accurate
model evaluation. The labeled data were used in training different machine
learning models like VGG16, InceptionV3, VGG19, and a CNN custom
model. The CNN custom model performed better among the selected
techniques and gave a classification accuracy of 90%. The findings suggest
that the custom CNN model can classify the banana types into four
categories, which is a great approach to automating the identification
process. Nonetheless, there are limitations, particularly the need for a bigger
dataset size to cater to different conditions. These limitations could be
mitigated in future studies by adding more banana types and a variety of data
in the training of the models. From the findings of this research, it is obvious
that farmers will be able to enhance crop management by employing this
technique to boost agricultural productivity.
