<![CDATA[Self-supervised learning with the SimCLR method requires appropriate data augmentation as the foundation for contrastive representation learning. In the task of classifying oil palm fruit maturity, color is the main discriminative feature, so the standard RGB-based ColorJitter augmentation has the potential to distort critical color information. This study proposes an augmentation framework based on the HSV color space with eight parameter configurations (C1–C8) determined in a data-driven manner through Eta-squared (η²) analysis, designed to preserve the semantic integrity of maturity color while generating varied positive pairs. Experiments were conducted on 1,016 oil palm fruit images across three classes (Unripe, Ripe, Overripe) using ResNet-18 as the SimCLR encoder with a linear evaluation protocol. The results show that all HSV configurations outperform the Baseline (82.22%), with HSV_C1 being the best configuration, achieving a test accuracy of 94.44% and Macro F1-Score of 94.46% (+12.22 pp). The η² analysis confirms that the Hue component is the most discriminative feature between classes, so constraining it during augmentation proves essential for the quality of the resulting representations. The proposed framework consistently improves SimCLR performance in color-based classification domains with minimal labeled data requirements.]]>
Copyrights © 2026
