Yapay Zeka Hızlı Bakış & Çıkarımlar
AI InsightThese findings support capacitive sensing as a promising, inexpensive approach for contactless olive screening under practical postharvest operating conditions, although validation across larger samples and multiple cultivars is required.
📌 3 Maddede Bu Makale
Özet
Bruising during harvest and postharvest handling reduces the quality, market value, and storage life of table olives. This study proposes a low-cost, fully contactless method for distinguishing bruised olives from sound fruit using capacitive sensing and machine learning. The sensing unit combines a capacitance-to-digital converter with custom copper parallel-plate electrodes. Olives passed individually between the electrodes while capacitance was continuously recorded, producing a 2,000-sample time series containing 18 passages: nine sound and nine bruised olives. Samples were labelled as background, sound olive, or bruised olive. To address background dominance, a balanced dataset of 538 sliding windows was created, and 13 statistical and signal-based features were extracted. Evaluation followed a leakage-free protocol: complete olive passages were held out, overlapping training windows were purged, and scaling and window-length selection were performed only within training folds. For balanced three-class classification, Random Forest achieved 71.93% accuracy, 71.92% balanced accuracy, and a macro F1-score of 0.720, compared with a 33.3% chance level. Olive detection reached an AUC of 0.830, while bruise classification achieved 81.84% accuracy, 84.70% specificity, and an AUC of 0.849. Passage-level analysis, using one feature vector per fruit, correctly classified 17 of 18 olives (94.4%; 95% CI: 0.727–0.999) and detected all bruised samples, with an AUC of 0.988. A conventional random split produced 91.67% window-level accuracy, demonstrating approximately 20 percentage points of leakage-related inflation. These findings support capacitive sensing as a promising, inexpensive approach for contactless olive screening under practical postharvest operating conditions, although validation across larger samples and multiple cultivars is required.
Abstract
Bruising during harvest and postharvest handling reduces the quality, market value, and storage life of table olives. This study proposes a low-cost, fully contactless method for distinguishing bruised olives from sound fruit using capacitive sensing and machine learning. The sensing unit combines a capacitance-to-digital converter with custom copper parallel-plate electrodes. Olives passed individually between the electrodes while capacitance was continuously recorded, producing a 2,000-sample time series containing 18 passages: nine sound and nine bruised olives. Samples were labelled as background, sound olive, or bruised olive. To address background dominance, a balanced dataset of 538 sliding windows was created, and 13 statistical and signal-based features were extracted. Evaluation followed a leakage-free protocol: complete olive passages were held out, overlapping training windows were purged, and scaling and window-length selection were performed only within training folds. For balanced three-class classification, Random Forest achieved 71.93% accuracy, 71.92% balanced accuracy, and a macro F1-score of 0.720, compared with a 33.3% chance level. Olive detection reached an AUC of 0.830, while bruise classification achieved 81.84% accuracy, 84.70% specificity, and an AUC of 0.849. Passage-level analysis, using one feature vector per fruit, correctly classified 17 of 18 olives (94.4%; 95% CI: 0.727–0.999) and detected all bruised samples, with an AUC of 0.988. A conventional random split produced 91.67% window-level accuracy, demonstrating approximately 20 percentage points of leakage-related inflation. These findings support capacitive sensing as a promising, inexpensive approach for contactless olive screening under practical postharvest operating conditions, although validation across larger samples and multiple cultivars is required.