Scholarlink

<p>Automated grading of date fruit is an important step in intelligent agriculture and in consistent post-harvest quality control for both local and export markets. This paper presents a feature-level multimodal deep learning framework for Algerian date fruit quality assessment and frames classification as a sensor-driven agricultural AI problem, in which reliable characterization requires integrating heterogeneous measurements rather than relying on RGB appearance alone. Two experimental configurations are investigated: (i) Scenario&#xa0;I (Multi-view RGB), which uses four RGB views per fruit sample, and (ii) Scenario&#xa0;II (Multimodal fusion), which augments the multi-view RGB setup with thermal imaging and a weight measurement for each fruit. Experiments on a locally collected dataset of 853 samples covering eight quality grades from two cultivars (<i>Deglet Nour</i> and <i>Mech-Degla</i>) sourced from an industrial sorting facility show that multimodal fusion improves performance across all evaluated architectures. For the best-performing custom CNN, validation accuracy increases from 80.7% in Scenario&#xa0;I to 89.5% in Scenario&#xa0;II, while macro-F1 improves from 77.4% to 87.6%. These improvements indicate more balanced class discrimination under class imbalance and fine-grained classification conditions. Analyses of learning curves and confusion matrices further show that thermal and physical indicators provide complementary information that reduces inter-class confusion, particularly for visually similar quality classes. The results highlight the practical value of sensor-driven multimodal learning for robust date fruit quality assessment under conditions closer to industrial classification practice, beyond unimodal vision-based pipelines.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Feature-level multimodal deep learning for date fruit quality assessment

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Ibtissam Boumaraf,
  • Abdelhamid Djeffal,
  • Abdelmalik Taleb-Ahmed

摘要

Automated grading of date fruit is an important step in intelligent agriculture and in consistent post-harvest quality control for both local and export markets. This paper presents a feature-level multimodal deep learning framework for Algerian date fruit quality assessment and frames classification as a sensor-driven agricultural AI problem, in which reliable characterization requires integrating heterogeneous measurements rather than relying on RGB appearance alone. Two experimental configurations are investigated: (i) Scenario I (Multi-view RGB), which uses four RGB views per fruit sample, and (ii) Scenario II (Multimodal fusion), which augments the multi-view RGB setup with thermal imaging and a weight measurement for each fruit. Experiments on a locally collected dataset of 853 samples covering eight quality grades from two cultivars (Deglet Nour and Mech-Degla) sourced from an industrial sorting facility show that multimodal fusion improves performance across all evaluated architectures. For the best-performing custom CNN, validation accuracy increases from 80.7% in Scenario I to 89.5% in Scenario II, while macro-F1 improves from 77.4% to 87.6%. These improvements indicate more balanced class discrimination under class imbalance and fine-grained classification conditions. Analyses of learning curves and confusion matrices further show that thermal and physical indicators provide complementary information that reduces inter-class confusion, particularly for visually similar quality classes. The results highlight the practical value of sensor-driven multimodal learning for robust date fruit quality assessment under conditions closer to industrial classification practice, beyond unimodal vision-based pipelines.