| 摘要: |
| 为提高温室环境最优控制中生菜信息在线反馈精度，通过群体图像识别研究生菜鲜重估算方法；通过生菜群体图像和单株图像，研究群体估算时误差正负相消对整体误差的改善作用，评估生菜遮挡问题对估算精度的影响，并研究能否通过改进深度学习的损失函数以实现对估算精度的进一步提高。结果表明：1）与不存在遮挡问题的单株图像生菜鲜重估算结果相比，基于群体图像裁剪的生菜鲜重估算决定系数（R2）低0.010 8， 归一化均方根误差（NRMSE）高2.69%，平均绝对百分误差（MAPE）低2.36%，虽然估算精度略低，但是生菜群体的遮挡问题更能反映生产实际。2）群体估算虽然存在遮挡问题导致裁剪不完整，但根据误差正负相消原理，相比没有遮挡的单株估算结果MAPE仍然低3.49%，因此更适用于生菜产量信息反馈。3）基于更优化MAPE的损失函数平均平方百分误差（MSPE），可以进一步降低群体估算的MAPE至8.46%，满足“软测量”对估算精度的需求。考虑到温室生菜的实际生产情况，群体估算更适合用于温室环境最优控制中生菜产量信息的在线反馈，通过深度学习等方法的优化，可以将生菜产量的估算误差降低至10%以内。 |
| 关键词: 群体估算 生菜鲜重 图像识别 深度学习 |
| DOI：10.11841/j.issn.1007-4333.2024.04.15 |
| 投稿时间：2023-09-25 |
| 基金项目:山东省重点研发计划项目（2022CXGC020708）；国家自然科学基金项目（32371998，U20A2020）；现代农业产业技术体系项目（CARS-23-D02）；北京市设施蔬菜创新团队项目（BAIC01-2023） |
|
| Image recognition of lettuce fresh weight through group estimation |
|
XU Dan1, LI Shuoguo1, CHEN Jingjing1, CUI Tingyuan2, ZHANG Yi2, MA Juncheng1*
|
| (1.College of Water Resources and Civil Engineering, China Agricultural University, Beijing 100083, China;2.Institute of Environment and Sustainable Development in Agriculture, Chinese Academy of Agriculture Sciences, Beijing 100081, China) |
| Abstract: |
| The lettuce fresh weight estimation through group images was researched to increase the accuracy of online feedback of lettuce information in optimal control of greenhouse climate. Based on the group and single lettuce images， the decrease of over error was investigated through cancellation of positive and negative errors based on group estimation. The influence of occlusion on estimation accuracy was quantified. The improvement of estimation accuracy was realized by introducing a novel loss function with the method of deep learning. The results showed that： 1）Compared with the estimation results of lettuce images in single plant images without occlusion， the determination coefficient（R2） of lettuce fresh weight estimation based on group estimation cropping decreased 0.010 8， the normalized root mean squared error（NRMSE） increased 2.69%， and the mean absolute percentage error（MAPE） decreased 2.36%. Although the estimation was slightly lower， the occlusion of the lettuce group can be close to real production. 2）Although there were occlusion issues in group estimation that led to incomplete cropping， the MAPE in group estimation with the cancellation of positive and negative errors was still 3.49% lower than that in single-lettuce estimation. Therefore， group estimation was a better way of providing feedback on lettuce yield. 3）Based on a more optimized MAPE， mean squared percentage error（MSPE） of the loss function can be further decreased to 8.46% which satisfied the requirements on the estimation accuracy of a “soft sensor”. Considering the reality of greenhouse lettuce production， group estimation is a better way to provide online feedback on lettuce yield to the optimal control of the greenhouse climate. Through optimization methods such as deep learning， the estimation error of lettuce yield can be decreased to within 10%. |
| Key words: group estimation lettuce fresh weight image recognition deep learning |
