A team of scientists from Pakistan has developed an AI-based method to accurately determine the sweetness of native citrus fruits, in a recent research published in Nature, a prestigious research journal.
The team, led by Dr. Ayesha Zeb from the National Centre of Robotics and Automation at NUST, achieved over 80% accuracy in predicting fruit sweetness without causing any damage to the fruit.
The researchers selected 92 samples from a farm in the Chakwal district for their experiment. They used a handheld spectrometer to capture patterns of light bouncing off the fruits’ skin and analyzed the fruit samples using near-infrared spectroscopy. Out of the 92 fruits, 64 were used for calibration and 28 were used for prediction.
What makes this unique is the team’s use of NIR spectroscopy to model the sweetness of local fruits. They integrated AI algorithms to directly classify the sweetness of oranges, resulting in improved accuracy.
Traditionally, fruit sweetness is determined through chemical and sensory testing. To develop their AI model, the team collected reference values for Brix, TA, and fruit sweetness by peeling off samples from the marked areas used for spectroscopy. They also had human volunteers taste the fruits and categorize them as flat, sweet, or very sweet.
Using the obtained spectra, reference values, and sweetness labels, the team trained their AI algorithm using a total of 128 samples. To evaluate its accuracy, they tested it with data from 48 new fruits.
The results were impressive – the AI model accurately predicted Brix, TA, and overall sweetness values and outperformed traditional methods in sweetness prediction. The model achieved an overall accuracy rate of 81 percent for identifying sweet, mixed, and acidic tastes.
This breakthrough has significant implications for the citrus industry in assessing fruit quality. This method has the potential to streamline and enhance the assessment of citrus fruits. With Pakistan being a major global producer of citrus fruits, this advancement holds great potential.