Song Yufei, Liu Aoran, Meng Xi, Liu Zhiguo, Liu Ping, Zhen Xinshuo
College of Horticulture, Hebei Agricultural University, Baoding, China.
School of College of Computer and Cyber Security, Hebei Normal University,Hebei Provincial Key Laboratory of Network & Information Security, Hebei Provincial Engineering Research Center for Supply Chain Big Data Analytics & Data Security, Shijiazhuang, China.
NPJ Sci Food. 2025 Aug 25;9(1):187. doi: 10.1038/s41538-025-00551-3.
Accurate, non-destructive classification of winter jujube maturity is critical for quality control and intelligent harvesting. This study proposes a dual-stream attention-fused residual network (DSAF-ResNet) combining hyperspectral and GLCM-based texture features at the feature level. The multimodal fusion significantly improved classification performance, with ResNet34 achieving 92.27% test accuracy under fused inputs. The DSAF-ResNet, integrating RepVGGBlock, SimAM attention, and a dual-stream architecture, achieved 98.61% training accuracy and 97.24% test accuracy, with 97.31% precision and 97.24% recall. Ablation experiments confirmed the contribution of each module. DSAF-ResNet demonstrated excellent generalization, stability, and robustness in distinguishing subtle maturity differences, even under class imbalance. This work provides an effective, scalable framework for non-destructive fruit maturity classification, advancing intelligent agricultural practices and supporting precision agriculture applications.
冬枣成熟度的准确、无损分类对于质量控制和智能采收至关重要。本研究提出了一种双流注意力融合残差网络（DSAF-ResNet），该网络在特征层面结合了高光谱和基于灰度共生矩阵的纹理特征。多模态融合显著提高了分类性能，在融合输入下ResNet34的测试准确率达到了92.27%。集成了RepVGGBlock、SimAM注意力和双流架构的DSAF-ResNet训练准确率达到98.61%，测试准确率达到97.24%，精确率为97.31%，召回率为97.24%。消融实验证实了每个模块的贡献。即使在类别不平衡的情况下，DSAF-ResNet在区分细微的成熟度差异方面也表现出了出色的泛化性、稳定性和鲁棒性。这项工作为水果成熟度的无损分类提供了一个有效、可扩展的框架，推动了智能农业实践，并支持精准农业应用。
