Scholarlink

<p>Saffron adulteration poses a threat to consumer confidence and the economic viability of Kashmir Valley saffron, a high-quality geographical indication product. Conventional approaches to saffron adulteration are subjective, time-consuming, and not scalable. This paper proposes SIPSNet, an explainable, statistically validated deep ensemble learning framework for effective and interpretable saffron adulteration detection. SIPSNet combines ResNet-50, VGG19, and InceptionV3 architectures using majority voting and is trained on SIPSNet, a newly constructed dataset of 1,200 high-resolution images of Kashmir Valley saffron. We use Grad-CAM for interpretability and McNemar’s test for statistical validation. SIPSNet obtains a testing accuracy of 98.97%, an F-score of 99.30%, and a validation loss of 0.005, which outperforms all individual models and existing approaches. Grad-CAM analysis verifies the model’s attention to important morphological attributes (stigma shape, colour homogeneity, and filament integrity). McNemar’s test validates the statistical significance of improvement over baseline models (<i>p</i> &lt; 0.05). SIPSNet performs well on our dataset, providing a quick, non-destructive, and useful method for safeguarding GI heritage and certifying saffron.</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

SIPSNet: a statistically validated and explainable deep ensemble framework for saffron adulteration detection

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Suhail Manzoor,
  • Syed Nisar Hussain Bukhari

摘要

Saffron adulteration poses a threat to consumer confidence and the economic viability of Kashmir Valley saffron, a high-quality geographical indication product. Conventional approaches to saffron adulteration are subjective, time-consuming, and not scalable. This paper proposes SIPSNet, an explainable, statistically validated deep ensemble learning framework for effective and interpretable saffron adulteration detection. SIPSNet combines ResNet-50, VGG19, and InceptionV3 architectures using majority voting and is trained on SIPSNet, a newly constructed dataset of 1,200 high-resolution images of Kashmir Valley saffron. We use Grad-CAM for interpretability and McNemar’s test for statistical validation. SIPSNet obtains a testing accuracy of 98.97%, an F-score of 99.30%, and a validation loss of 0.005, which outperforms all individual models and existing approaches. Grad-CAM analysis verifies the model’s attention to important morphological attributes (stigma shape, colour homogeneity, and filament integrity). McNemar’s test validates the statistical significance of improvement over baseline models (p < 0.05). SIPSNet performs well on our dataset, providing a quick, non-destructive, and useful method for safeguarding GI heritage and certifying saffron.