多光谱图像信息复合的作物病害稀疏特征识别方法

Crop diseases information was lost caused by the single identification image or spectrum, because many disease identification methods contain only the external disease characteristics or the lack of spatial information. In this project, sparse feature mining method for crop diseases with temporal and spatial information is studied, in order to reduce the influence of redundant information on the spectral and image sparse features of different diseases in time axis. A spectral band selection algorithm for crop diseases based on instability index, and calculate the inter class instability index of typical diseases corresponding bands, and then achieve the best corresponding spectral bands and different types of disease by optimal threshold of spectral curve and disease types. The intelligent diagnosis of crop diseases by environmental/visible/spectral synchronization is studied, the sparse reconstruction for leaf color, shape, texture and spectral characteristics, and then classification and combination of the external characteristics and spectral of the typical diseases, finally, the internal and external characteristics of crop diseases can be accurately identified. Through project research, provide the theoretical basis for control and cut off the spread of crop diseases, crop precision and digital management.

传统作物病害诊断方法一般采用图像或光谱技术进行单一检测，缺少目标对象相关的土壤、营养和环境等基础数据，仅从作物外部病害特征角度识别，识别准确性和稳定性较低。项目研究时空信息驱动的作物病害稀疏特征挖掘方法，融合不同生育期作物生长机理与环境胁迫模型，差异化提取病害光谱与图像，对多元光谱与纹理特征共生矩阵进行稀疏特征深度挖掘，建立光谱反射率与作物表征图像耦合关系消除内外部变量冗余；研究高维稀疏分解聚类的类间不稳定光谱波段优选模型，建立健康/患病叶片的光谱曲线与染病种类最优阈值，实现光谱波段与病害种类最大信息匹配；研究群体智能优化的作物病害环境/图像/光谱同步诊断方法，对典型病害叶片颜色、形状、纹理等外部特征与光谱特征波长稀疏重构、分类、组合，实现作物病害内外部综合特征的精确识别。通过项目研究，形成多光谱图像信息复合的作物病害稀疏特征识别方法，为精准农业病害准确预测与防治提供理论基础。

作物病害检测方法中，多数只是进行图像或光谱技术的单一检测，只能根据作物外部病害特征进行识别或缺少目标对象的空间信息，造成病害特征丢失，影响识别的稳定性。项目针对作物病害数据集数量较小，发病早期图像识别率低的问题，改进作物病害图像识别模型，研究新的注意力模块的浅层识别网络，融合结构化作物生长环境参数与非结构化图像等多结构数据，实现对在不同环境参数下作物易发典型病害的智能诊断分析；挖掘光谱反射曲线的变化规律，研究基于时间序列的病害分析方法，学习健康叶片与染病叶片在不同观测维度间的差异信息表达，实现作物病害早期检测；将深度学习模型作为数据融合过程中的特征提取分类模型，研究基于多光谱图像信息融合的作物病害稀疏特征识别方法，从不同维度挖掘与病害相关的光谱信息和图像信息，实现作物病害的快速、准确、无损诊断。通过项目研究，为提升病害监测实时性和准确性，控制和切断作物病害大面积蔓延，开展作物精准化和数字化管理提供理论基础。发表论文33篇，其中SCI检索论文15篇，EI检索期刊6篇，北大中文核心期刊5篇，其他论文7篇，申请发明专利3项。