Zhang Jing, Zhou Hao, Liu Kunyu, Xu Yuguang
College of Artificial Intelligence & Computer Science, Xi'an University of Science and Technology, Xi'an 710600, China.
School of Economics and Management, Xidian University, Xi'an 710126, China.
Sensors (Basel). 2025 Apr 12;25(8):2432. doi: 10.3390/s25082432.
The outbreak of cassava diseases poses a serious threat to agricultural economic security and food production systems in tropical regions. Traditional manual monitoring methods are limited by efficiency bottlenecks and insufficient spatial coverage. Although low-altitude drone technology offers advantages such as high resolution and strong timeliness, it faces dual challenges in the field of disease identification, such as complex background interference and irregular disease morphology. To address these issues, this study proposes an intelligent classification method for cassava diseases based on drone imagery and an ED-Swin Transformer. Firstly, we introduced the EMAGE (Efficient Multi-Scale Attention with Grouping and Expansion) module, which integrates the global distribution features and local texture details of diseased leaves in drone imagery through a multi-scale grouped attention mechanism, effectively mitigating the interference of complex background noise on feature extraction. Secondly, the DASPP (Deformable Atrous Spatial Pyramid Pooling) module was designed to use deformable atrous convolution to adaptively match the irregular boundaries of diseased areas, enhancing the model's robustness to morphological variations caused by angles and occlusions in low-altitude drone photography. The results show that the ED-Swin Transformer model achieved excellent performance across five evaluation metrics, with scores of 94.32%, 94.56%, 98.56%, 89.22%, and 96.52%, representing improvements of 1.28%, 2.32%, 0.38%, 3.12%, and 1.4%, respectively. These experiments demonstrate the superior performance of the ED-Swin Transformer model in cassava classification networks.
木薯病害的爆发对热带地区的农业经济安全和粮食生产系统构成了严重威胁。传统的人工监测方法受到效率瓶颈和空间覆盖不足的限制。尽管低空无人机技术具有高分辨率和强时效性等优势，但在病害识别领域面临复杂背景干扰和病害形态不规则等双重挑战。为解决这些问题，本研究提出了一种基于无人机图像和ED-Swin Transformer的木薯病害智能分类方法。首先，我们引入了EMAGE（具有分组和扩展的高效多尺度注意力）模块，该模块通过多尺度分组注意力机制整合无人机图像中病叶的全局分布特征和局部纹理细节，有效减轻复杂背景噪声对特征提取的干扰。其次，设计了DASPP（可变形空洞空间金字塔池化）模块，使用可变形空洞卷积自适应匹配病害区域的不规则边界，增强模型对低空无人机拍摄中角度和遮挡引起的形态变化的鲁棒性。结果表明，ED-Swin Transformer模型在五个评估指标上均取得了优异成绩，得分分别为94.32%、94.56%、98.56%、89.22%和96.52%，分别提高了1.28%、2.32%、0.38%、3.12%和1.4%。这些实验证明了ED-Swin Transformer模型在木薯分类网络中的卓越性能。
