Non-destructive assessment of watermelon internal quality using image processing, machine learning, and smartphone-based decision support

This is the authors' abstract. We don't add key points for this paper.

Watermelon ( Citrullus lanatus ) is a widely consumed fruit worldwide; however, varietal differences are poorly characterized and largely unrecognized by consumers. Consequently, assessment of internal quality remains challenging, as it cannot be reliably inferred from external attributes alone. This study aimed to compare the functional properties of three watermelon varieties (Sweet Baby, Rocky 475, Kinaree 457) and develop a machine learning-based mobile app to predict their internal quality attributes. The functional properties of the selected varieties were analyzed using Total Phenolic Content (TPC), 2,2-Diphenyl-1-picrylhydrazyl (DPPH) free radical scavenging activity, and Ferric Reducing Antioxidant Power (FRAP). External features extracted from watermelon images were used to train five ML models to predict watermelon variety, sweetness, flesh color, and firmness. In addition, a survey was conducted to assess consumer satisfaction with market-available watermelons. Survey results showed that over 50% of respondents were dissatisfied with the sweetness and flesh color of watermelons available in the market. Biochemical analyses revealed significant (p<0.05) differences in bioactive compounds and antioxidative activities between watermelon varieties. Kinaree 457 showed the highest TPC (5.07 mg GAE/100g FW) and DPPH activity (7.47 mg TE/100g FW), while Sweet Baby exhibited the highest FRAP activity (5.96 mg Fe²⁺ Eq/100g FW). Correlation analysis (r>0.8) showed strong associations between internal quality traits and features such as fruit weight, pixel area, and entropy. The Random Forest model achieved the highest mean accuracy (95.98%), precision (96.86%), recall (93.97%), and F1-score (95.22%) across the four predicted attributes and was subsequently integrated into the mobile application. The study results demonstrate the application’s potential as a scientific tool to support informed consumer decision-making.