Real-time Optimization using ANN/Deep Convolutional Neural Network for Lowbush Blueberry Harvesting

到2050年，农业将在养活全球预计超过90亿的人口方面发挥至关重要的作用。然而，这个行业面临的令人担忧的情况是，世界上农场经营者的总数在不断下降。在加拿大，由于农业劳动力老龄化，在过去的二十年里，它下降了25%。低丛蓝莓是加拿大主要的园艺作物。在过去十年中，加拿大蓝莓的总种植面积增加了61%。劳动力短缺也是低丛蓝莓产业不可避免的问题。目前，80%以上的低丛蓝莓是由机械收获机收获的。机械收割机真正依靠操作人员的技能和经验来提高果实的回收率和质量，同时减少对收割机的损害。然而，由于劳动力老龄化和低丛蓝莓的采收窗口极短（大约3到4周），低丛蓝莓产业很难找到足够有经验的采收工人。因此，收获机自动化是矮丛蓝莓产业的迫切需要。由于许多因素相互关联，以前使用的多重传感和数学优化不足以预测最佳收获设置。人工神经网络（Artificial Neural Network， ANN）广泛应用于农业领域，但以往的方法都不是实时的现场解决方案。本研究提出了一种基于短期目标的低灌木采伐优化实时优化方法；(i)开发传感器融合系统，（ii）开发基于硬件的快速图像处理系统的架构和方法，以及（iii）利用人工神经网络/深度卷积神经网络进行实时建模。在此基础上，未来将利用神经网络建模程序和实时现场传感数据推导出最佳参数，从而实现可交付的嵌入式系统。人工神经网络/深度卷积神经网络实时嵌入式系统是农业自动化和机器人技术的新时代。收获的优化具有解决劳动力短缺危机的巨大潜力，因为它将用于收获机自动化，并将增加低丛蓝莓产业的可持续性。收获效率提高5%，每年将为加拿大低丛蓝莓产业带来5500万美元的收入。收获的优化将为生物系统自动化研究计划提供良好的基础，因为它可以很容易地转移到其他种植系统和其他农业部门，如使用人工神经网络/深度卷积神经网络进行动物行为分析。此外，该计划训练有素的HPQs将在不同的农业部门工作，为高技能的加拿大农业自动化和机器人人才提供良好的基础。

The agricultural industry will play a vital role in feeding over 9 billion predicted population on the globe by 2050. However, the alarming situation facing this industry is the total number of farm operators in the world is constantly declining. In Canada, it declined by 25% in the last two decades mainly due to an aging agricultural labor force. Lowbush blueberry is a dominant horticultural crop in Canada. The total acreage of a blueberry in Canada was increased by 61% in the last decade. Labour shortage is also an inevitable problem of lowbush blueberry industry. Currently, more than 80% of lowbush blueberry is harvested by mechanical harvester. The mechanical harvester really relies on operator skills and experience for better fruit recovery and quality with less damage to the harvester. However, due to the aging labor force and an extremely short harvesting window for lowbush blueberry (around 3 to 4 weeks), the lowbush blueberry industry has a difficulty to find enough experienced harvester operators. Therefore, harvester automation is an urgent need for the lowbush blueberry industry.Previously used multiple-sensing and mathematical optimization is not enough to predict the optimal harvesting set-up because so many factors are interrelated. Artificial Neural Network (ANN) was widely used for many agricultural applications however, none of the previous approaches were real-time in field solutions. This research proposes a real-time optimization for optimal lowbush harvesting by accomplishing short term objectives; (i) to develop a sensor fusion system, (ii) to develop an architecture and methodology for a hardware based fast image processing system, and (iii) a real-time modeling utilizing an ANN/Deep Convolutional Neural Network. Based on this research, a deliverable embedded system will be made in the future deriving the optimum parameters using the neural network modeling program and real-time field sensing data.The real-time embedded system of ANN/ Deep Convolutional Neural Network is a new era of agricultural automation and robotics. The optimization of the harvesting has huge potential to solve labour shortage crisis as it will be utilized for harvester automation and will increase the sustainability of the lowbush blueberry industry. Five percent increase in harvesting efficiency would result in $55 million of revenue to Canadian lowbush blueberry industry per year. The optimization of the harvesting will serve a good foundation for the Bio-systems Automation Research Program as it can be easily transferred to other cropping systems and other agricultural sectors like animal behavior analysis using ANN/ Deep Convolutional Neural Network. Moreover well trained HPQs from this program will work within different agricultural sectors, providing a good foundation of highly skilled Canadian agricultural automation and robotics personnel.