Crop Stress Management using Multi-source Data Fusion

精准农业（PA）使农业投入合理化，减轻农业实践对环境的不利影响，同时提高农业生产力和盈利能力。PA要求准确检测、分类、量化和绘制生物（杂草、害虫）和非生物（水、养分缺乏、化学品、热）胁迫，作为特定地点胁迫管理的主要信息。土壤类型、含水量、地表地形和遥感卫星数据补充了这一信息，以便部署有效的应力控制。该研究计划旨在通过应用计算机视觉（CV）和机器学习（ML）技术进行特定地点的压力管理来改善PA的整体决策支持系统。在PA中基于CV和ML的应用存在几个公开的挑战，例如（1）代表用于应力管理的真实的现场条件的公开可用的图像数据集的稀缺;（2）对这样的数据进行极其困难和繁琐的预处理;（3）基于单一类型的方法。（主要是图像）数据的目标问题;（4）未解决的闭塞和重叠问题;和（5）缺乏整体的方法来管理压力。该研究计划的长期目标是应用CV和ML技术为农业部门开发具有成本效益，可持续，环境友好和社会可接受的技术解决方案。短期目标将在未来五年内解决上述差距，即：（1）开发用于农业应用的半自动图像标记工具，（2）基于多源图像数据创建新的田间植被指数，以改进对作物发育的时间监测，（3）识别、分类、量化、并通过将ML技术应用于高分辨率图像、土壤数据、水分数据和天气数据等多光谱数据来监测生物和非生物胁迫（4）通过使用ML技术来检测，定量和动态阈值的叶面损害所造成的跳甲。该研究计划将为做出明智的决策提供基础，并将有助于经济，环境友好和社会可行的农业实践。这将对加拿大经济和世界粮食供应产生积极影响。四名高素质的人员将接受ML，CV和PA领域的培训。该研究计划与联邦人工智能创新议程（泛加拿大人工智能战略）和里贾纳大学的战略研究计划（数字未来集群和地面安全倡议）保持一致。

Precision Agriculture (PA) rationalizes farm inputs and mitigates the adverse impact of agricultural practices on the environment while increasing farm productivity and profitability. PA requires accurate detection, classification, quantification and mapping of biotic (weeds, pests) and abiotic (water, nutrient deficiency, chemicals, heat) stresses as the primary piece of information in site-specific stress management. Soil type, water content, surface topography and remote sensing satellite data complement this information for deploying effective stress control. This research program is designed to improve overall decision support systems in PA through application of Computer Vision (CV) and Machine Learning (ML) techniques for site-specific stress management. There are several open challenges with CV and ML-based applications in PA such as (1) the scarcity of publicly available image datasets representing real field conditions for stress management; (2) extremely difficult and tedious preprocessing of such data; (3) methods based on single type (primarily images) data for target problems; (4) unaddressed occlusion and overlap problem; and (5) lack of holistic approach for stress management. The long-term objective of this research program is to apply CV and ML techniques to develop technology solutions for the agriculture sector that are cost-effective, sustainable, environment friendly, and socially acceptable. The short-term objectives, which will address the aforementioned gaps over the next five years are to (1) develop a semi-automated image labelling tool for agricultural applications, (2) create novel field vegetation indices based on multi-source image data for improved temporal monitoring of crop development, (3) identify, classify, quantify, and monitor biotic and abiotic stresses by applying ML techniques on multisource data like high-resolution imagery, soil data, moisture data and weather data (4) develop a decision support system for canola crop management by using ML techniques to detect, quantify and dynamically threshold foliation damage caused by flea beetle. This research program will provide the basis for making informed decisions and will help in economical, environment-friendly, and socially viable agricultural practices. This will have a positive impact on the Canadian economy and world food supply. Four highly qualified personnel will be trained in the areas of ML, CV, and PA. The research program is aligned with the federal innovation agenda on artificial intelligence (Pan-Canadian Artificial Intelligence Strategy) and the strategic research plan of the University of Regina (digital future cluster and ground security initiative).