The AI-based approach makes analytical equipment like high-performance liquid chromatography redundant.
Researchers at the School of Biotechnology, Jawaharlal Nehru University, New Delhi have developed an Artificial Intelligence-based tool to predict the level of metabolite content in neem fruits as either low or high based on images of neem leaves and fruits. The AI-based approach makes the use of expensive and cumbersome analytical equipment like high-performance liquid chromatography (HPLC) redundant. The classification of fruits can be carried out in the field even by untrained people. The results have been posted on a preprint server; preprints are not peer-reviewed.
The key metabolite of neem fruits is azadirachtin, which is used as a biopesticide. However, azadirachtin content in the seed kernel varies from one tree to another. Since the fruits with different concentrations of azadirachtin are mixed and bagged together in the field, the final azadirachtin content in each sack becomes low. Currently, there is no easy tool to evaluate the azadirachtin content level (high or low) in neem fruits inexpensively. Therefore, only a few fruits from each sack with varying azadirachtin content or from a single batch of multiple sacks are taken to determine the azadirachtin content using HPLC.
The team led by Dr. Binay Panda from JNU collected 1,045 neem leaves and fruits from trees across India and imaged them and also determined the content of five metabolites — azadirachtin, salannin, deacetylsalannin, nimbin, and nimbolide — from neem fruits using HPLC. Synthetic image augmentation while training using the deep learning frameworks ensured the sufficiency of images, which served the purpose during training. “We used images of leaves and fruits and with the corresponding HPLC-measured metabolite values within different deep learning-based frameworks to test for accuracy,” says Dr. Panda. “Eighty percent of the data comprised of images and their corresponding metabolite values were first used for training, followed by 10% for model validation and the remaining 10% for model testing.”
The sensitivity
The sensitivity of the AI model to correctly determine and classify the fruits into low or high metabolite content was 83% and 82% for low and high azadirachtin classes, respectively when only the azadirachtin model was used. Using the same model, the specificity was 79% and 85% for low and high azadirachtin classes, respectively. “But when all five metabolites for the leaf and fruit images were considered, the sensitivity in predicting low and high classes was boosted by about 9% and 6%, respectively. Using the multi-analyte model, the specificity was boosted to 100% for both low and high classes,” Dr. Panda says.
“Our model makes it possible to use fruit and leaf images alone to predict the azadirachtin content class in fruits without having to use HPLC.”
Dr. Panda is not sure what parameters in the images of leaves and fruits were likely used by the model to determine the metabolite content in the fruits and classify them as either high or low classes. “There must be specific fruit and leaf features that are governed by metabolites, which change depending on the concentration of the metabolites,” he says.
The researchers have built an Android App called Fruit-in-Sight using the best predictive model.
“All that the fruit pickers need to do is take a picture of the leaf and fruit of a neem tree, and the App will instantly tell them whether to pick or not pick the fruits from the tree based on its azadirachtin content class prediction,” Dr. Panda says.
“This will help empower villagers to pick fruits from suitable trees with high azadirachtin content.”
He adds that the App is a simple yet powerful enough tool to help boost the efficiency of the neem industry without any additional cost, time, or effort.
