An Ordinal Dataset for Ripeness Level Classification in Oil Palm Fruit Quality Grading

This study hypothesizes that an ordinal dataset tailored for ripeness classification can enhance automated quality grading in the palm oil industry. The dataset, comprising 4,728 high-resolution images of oil palm fruits categorized into five ripeness levels (Immature, Partially Ripe, Fully Ripe, Overripe, and Decayed), was collected in real-world conditions from Central Kalimantan, Indonesia, using diverse devices and capturing environmental variability such as lighting, poses, and natural backgrounds. Defined with expert input, the categories reflect biological and economic relevance, addressing challenges like imbalanced data distributions typical in agriculture. Notable findings include the dataset’s authenticity, which mirrors real-world agricultural conditions and enables the development of robust machine learning models. With stratified splits for training, validation, and testing, the dataset facilitates benchmarking while supporting techniques like weighted loss functions to address imbalances. Its diversity and realistic complexity make it a valuable resource for advancing ordinal regression and automated grading systems, driving efficiency and sustainability in palm oil production.

本研究假设，针对成熟度分类量身定制的一种序数数据集能够提升棕榈油行业自动化质量分级。该数据集包含4,728张高分辨率油棕果图像，这些图像被划分为五个成熟度等级（未成熟、部分成熟、完全成熟、过熟和腐烂），数据采集于印度尼西亚中加里曼丹的真实环境，采用多种设备，并捕捉了光照、姿态和自然背景等环境变量的变化。在专家意见的基础上，这些类别反映了生物学和经济学的重要性，并解决了农业中常见的数据分布不平衡等挑战。显著发现包括数据集的真实性，其反映了真实的农业环境，并有助于开发稳健的机器学习模型。通过分层划分用于训练、验证和测试，该数据集促进了基准测试，同时支持诸如加权损失函数等技术，以解决不平衡问题。其多样性和现实复杂性使其成为推进序数回归和自动化分级系统的重要资源，推动棕榈油生产的效率和可持续性。