Real-time vision-based spot spraying development for high efficiency and precision weed management

杂草通常通过在整个田地中均匀喷洒化学品来控制。然而，在这一方法中过度使用化学品增加了作物保护的成本，对环境和粮食安全造成了负面影响，阻碍了可持续农业发展。定点喷洒作为一种空间可变的杂草管理策略，仅针对田间杂草种类，以尽量减少化学品的使用。基于感测植被光学性质的商业上可获得的技术通常受到检测土壤背景上的杂草（即，裸土壤背景中的绿色检测）的限制，并且不适合于检测生长中的作物中的杂草。基于视觉的定点喷洒系统能够区分植被物种。基于视觉的点喷技术发展的关键之一是建立一个可靠、鲁棒的杂草/作物识别模型。传统上，基于视觉的杂草/作物鉴别模型的开发高度依赖于具有杂草和作物之间所定义的颜色、纹理和形态特征的先验知识的图像分析。但这可能无法推广到不同的作物领域与多种杂草物种。最近机器学习和计算机视觉技术的进步为在非结构化田间条件下开发一个强大而可靠的基于视觉的杂草/作物识别模型提供了新的机会。目前，训练深度机器学习模型需要大量标记的图像。在该项目中，为了弥补手动标记任务并加快开发过程，提出了一种新的管道，以基于传统图像分析和生成对抗网络（GAN）的组合生成逼真的合成图像。此外，该项目还将致力于解决目前基于视觉的点喷发展中的瓶颈问题，即识别模型的泛化能力和喷洒效率。具体而言，本项目的目标是（1）建立两个图像数据库（杂草和作物），允许通过开发的生成管道快速生成带注释的合成图像;（2）基于大量合成图像开发一个强大而轻量级的基于深度学习的检测模型，以及（3）将所研制的杂草检测系统集成到喷杆上进行定点喷洒，并在田间条件下验证了其用于高效定点喷洒（行驶速度> 8 km/h）的可行性。示范将在甜菜地展出。将从喷洒精度和效率方面进行评价。

Weeds are typically controlled by spraying chemicals uniformly across the whole field. However, the overuse of chemicals in this approach has increased the cost of crop protection and posed negative impacts on the environment and food security, which is a hindrance to sustainable agriculture development. Spot spraying, as a spatially variable weed management strategy, targets only weed species in fields to minimize the use of chemicals. Commercially available technologies based on sensing of vegetation optical properties are typically constrained by detecting weeds on a soil background (i.e. greenness detection in a bare soil background) and are not suitable to detect weeds among a growing crop. A vision-based spot spraying system enables discrimination between vegetation species. One of the key components for the vision-based spot spraying development is to build a reliable and robust weed/crop discrimination model. Traditionally, the development of a vision-based weed/crop discrimination model is highly relying on image analysis with prior knowledge of the defined colour, texture and morphology features between weed and crop. But this might fail to generalise over different crop fields with multiple weed species. The recent technology advancements in machine learning and computer vision have provided new opportunities to develop a robust and reliable vision-based weed/crop discrimination model under unstructured field conditions. Currently, training a deep machine learning model require large numbers of labelled images. In this project, in order to remedy the manual labelling task and speed up the development process, a novel pipeline is proposed to generate realistic synthetic images based on the combination of conventional image analysis and generative adversarial networks (GANs). Furthermore, the project will also focus on dealing with the current bottlenecks of vision-based spot spraying development with regards to discrimination model generalization ability and spraying efficiency. Specifically, the objectives of this project are to (1) build two image data libraries (weed and crop) that allow fast annotated synthetic images generation via the developed generation pipeline; to (2) develop a robust and lightweight deep learning-based detection model based on large numbers of synthetic images and to (3) integrate the developed weed detection system into a spray boom for spot spraying and demonstrate its feasibility for high efficiency spot spraying (driving speed > 8 km/h) under field conditions. The demonstration will be displayed at a sugar beet field. The evaluation will be made in terms of spraying accuracy and efficiency.