Application of Imaging Technology and Machine Learning to Assess the Behavior of Parthenocarpic and Non-Parthenocarpic Cucumber Cultivars Under Lacto-Fermentation

The behavior of parthenocarpic and non-parthenocarpic cucumber during lacto-fermentation may be different. The research material consisted of two parthenocarpic cucumber cultivars ‘Malika’ F1 and ‘Magellan’ F1 and two conventional non-parthenocarpic cucumber cultivars ‘Zefir’ and ‘Ikar’. Raw material was subjected to spontaneous lacto-fermentation for 56 days and changes in cucumber flesh were assessed after selected periods of the process using texture features from images acquired using a flatbed scanner. The machine learning models based on image textures were built to discriminate raw material (0 days) and samples lacto-fermented for 3, 7, 10, 14, 28, and 56 days. For parthenocarpic cucumbers, an average accuracy of up to 88.0% for a model built based on selected image textures using Cubic SVM was obtained for ‘Malika’ F₁ and 91.3% (Cubic SVM) for ‘Magellan’ F₁. Whereas in the case of non-parthenocarpic cultivars, an average classification accuracy of 95.4% (Medium Neural Network) was observed for ‘Zefir’ and 93.1% (Cubic SVM) for ‘Ikar’. The greatest differences between individual samples were found in the case of non-parthenocarpic cucumber. An accuracy of 100% was determined for raw material and samples after 3 and 56 days of lacto-fermentation for both non-parthenocarpic cultivars. In the case of parthenocarpic cucumber cultivars, only raw material was correctly distinguished from lacto-fermented samples in 100%. The developed approach can be used in practice to determine the effect of lacto-fermentation on cucumber flesh in an objective and non-destructive manner and to select the most desirable cultivars for this process.