基于SCBAM-BiGRU-ResNet50的水稻缺素识别方法

Rice Nutrient Deficiency Identification Method Based on SCBAM-BiGRU-ResNet50

摘要:

在水稻生长过程中，氮、磷、钾等营养元素缺失会导致叶片变黄、发育迟缓等症状，严重影响水稻生长质量和最终产量。针对传统卷积神经网络在缺素类别识别上泛化能力不足，而新型架构（如EfficientNetV2等）因偏重全局建模而在局部病变区域的特征提取能力不足等问题，本文提出了基于卷积神经网络、循环神经网络及注意力机制相结合的识别模型SCBAM-BiGRU-ResNet50。在ResNet50骨干网络中引入风格化卷积注意力机制模块（Style-based convolutional block attention module，SCBAM），通过增强局部特征表达能力，使模型更聚焦于缺素状态下典型特征。同时结合双向门控循环单元（Bidirectional gated recurrent unit，BiGRU），利用双向时序学习特性提升特征提取能力。试验结果表明，在水稻缺素特征识别任务中SCBAM-BiGRU-ResNet50准确率为96.97%，表现最佳、显著优于对照组（p<0.0001），同时保持了较小的波动范围，验证了本文提出的改进结构在特征表达与序列建模方面的有效性和稳定性。该模型通过融合注意力机制和双向时序建模，有效提升了局部特征响应和上下文理解能力，为当前水稻缺素诊断提供新的解决方案。

Abstract:

During the whole growth cycle of rice, insufficient supply of core essential elements, including nitrogen, phosphorus and potassium easily triggers typical abnormal physiological phenomena, such as leaf chlorosis, slow tillering and overall plant stunting, which not only deteriorate rice appearance quality and nutritional quality, but also cause substantial reduction in actual grain yield and restrict stable agricultural production.Aiming at the problems that traditional convolutional neural networks lacked sufficient generalization ability in identifying nutrient deficiency categories, and novel architectures such as EfficientNetV2 were biased toward global modeling and thus had insufficient feature extraction capability for local lesion areas, an intelligent identification model named SCBAM-BiGRU-ResNet50 was constructed by integrating convolutional structure, recurrent neural network and optimized attention mechanism.The proposed method embeded improved style-based convolutional block attention module into ResNet50 backbone network to strengthen local feature mining and guided the model to concentrate on distinguishable deficiency-related leaf traits.Simultaneously, the integration of bidirectional gated recurrent units (BiGRU) leveraged bidirectional temporal learning to further improve feature extraction.Experimental results demonstrated that SCBAM-BiGRU-ResNet50 achieved superior performance with the accuracy index of 96.97% in rice nutrient deficiency classification tasks, significantly outperforming benchmark models in accuracy (p<0.0001) while maintaining low variability.This hybrid network structure effectively boosted local feature sensitivity and deepened contextual feature correlation learning.The research result can provide an efficient, accurate and stable technical reference and innovative intelligent diagnosis scheme for rapid and precise identification of rice nutrient deficiency in modern precision agriculture.