基于空地多视角RGB影像协同的病害香蕉植株监测

Monitoring of diseased banana plants based on collaborative RGB images from air-ground multi-view

  • 摘要: 【目的】利用空地RGB影像探索病害香蕉植株精准监测方法,为构建面向大面积香蕉种植区域中发病香蕉高效精准监测的智能系统提供技术支撑。【方法】通过无人机和地面移动设备拍摄,建立空地多视角病害香蕉植株RGB影像数据集,综合注意力机制和深度可分离卷积策略开发出检测大面积种植区域中病害香蕉植株数量的轻量化深度学习模型,运用多视角目标校对计算实际病害香蕉植株数量及发病率信息。【结果】改进后的模型比YOLOv3、YOLOv4和YOLOv5中最优基础模型性能指标更优。在空地多视角病害香蕉植株目标检测试验中,该模型全类平均精确率(m AP)达94.95%。基于该模型完成了大面积种植区域病害香蕉植株的监测,且发现病害香蕉植株多位于曾遭受过病害影响而形成的秃斑一样的区域内或其边界上,表明病害香蕉植株具有传染性。此外,完成了规范化小型试验地块中处于缓苗和营养生长期的中蕉4号香蕉的病害信息统计,模型检测所得香蕉发病率误差为0.31%,需人工巡检校对的数量大幅下降82.98%~100.00%,且发现该品种香蕉也表现出典型的多年生作物生长特征。【建议】为解决大面积种植区域中病害香蕉植株精准监测难和检测模型泛化能力不足的问题,建议用空地RGB影像协同的轻量化病害香蕉植株监测方法提取香蕉作物病害信息;多方协作,探索适合数字化农业发展的香蕉种植模式;延伸产业链发展,提升香蕉种植的综合效益,促进区域优势产业发展。

     

    Abstract: 【Objective】The precise monitoring method of diseased banana plants was explored by air-ground RGB image, to provide technique support for building intelligent system for efficient and accurate monitoring of diseased banana for large-scale planting areas. 【Method】The air-ground multi-view diseased banana plants RGB image dataset were established by unmanned aerial vehicle and ground mobile devices. A lightweight deep learning model for detecting the number of diseased banana plants in a large-scale planting area has been developed by combining attention mechanism and deep separable convolution strategy. The actual number of diseased banana plants and the incidence rate information were calculated by the multi-view target calibration method. 【Result】The performance metrics of the improved model were compared and found to perform better than the optimal base model in the commonly used YOLOv3,YOLOv4 and YOLOv5.In the experiments of target detection of diseased banana plants with multi-view, the mean average precision(mAP) of the model reached 94.95%. Based on the improved model, the monitoring of diseased banana plants in large planting areas was completed, and it was found that most of the diseased banana plants were located in the area or boundary of the bald spots formed by the disease, which indicated that the diseased banana plants were infectious. In addition, disease information statistics were completed for Zhongjiao No.4 in the slow and nutritive growth stages in the standardized smallscale experimental plots. The error of banana incidence rate from the model detection was less than 0.31%, the number of manual inspections and proofreading was greatly reduced by 82.98% to 100.00%, and it was found that this banana variety also showed typical perennial crop growth characteristics. 【Suggestion】In order to solve the problems of difficult accurate monitoring of diseased banana plants in large planting areas and insufficient generalization ability of detection models, it is recommended to extract banana crop disease information by lightweight diseased banana plant monitoring method based on collaborative RGB images from air-ground multi-view; multi-party collaboration was recommended to explore banana planting models suitable for the development of digital agriculture; extend the development of the industrial chain, improve the comprehensive benefits of banana planting, and promote the development of regional advantageous industries.

     

/

返回文章
返回