Knowledge Distillation empowered Mobile Intelligence Solution for Sustainable Management of Crop Pests and Soil Health

在世界范围内，农作物受到无脊椎动物害虫的威胁，这些害虫造成摄食损害并传播植物病毒。高水平的虫害可导致高达80%的产量损失。虽然农药经常用于作物保护，但其不加控制的使用会造成土壤侵蚀和污染。害虫的可持续管理依赖于：准确识别现有害虫，了解可容忍的害虫损害程度，以及有效的害虫管理解决方案，以保持土壤健康。目前，英国还没有可持续管理小麦害虫的综合田间解决方案。随着杀虫剂使用限制的增加和更多抗药性害虫种群的出现，种植者正在寻求更可持续的害虫综合管理（IPM）实践。然而，有许多障碍限制了IPM原则的采用：准确识别无脊椎害虫很困难，需要分类培训，而种植者往往缺乏这种技能;目前的阈值在田间条件下几乎没有得到测试和验证，限制了种植者的信心，在最初的项目中，我们通过构建一个AI驱动的移动的害虫，开发了一个早期解决方案。用于识别小麦作物中害虫的检测解决方案（创新项目10002902）。结果可以进一步改进，以获得更大的成功，其好处如下：* 将小麦害虫检测模型扩展到马铃薯和油菜等以上作物。通过优化机制提高移动的手机中运行的模型的识别准确性和效率。*评估和验证重点农作物害虫的可接受害虫阈值。在这里，我们将提出一个后续项目，通过优化深度学习检测模型和知识蒸馏技术来改进我们的人工智能驱动的移动的害虫管理解决方案，以用于其他耕地作物的害虫。其结果将是一个增强的MPM解决方案，该解决方案提供：1）使用移动的设备快速检测和准确量化叶面害虫; 2）将害虫量化置于特定区域害虫耐受阈值的背景下; 3）提供经济阈值估计和害虫控制建议。它将建立在联盟现有资源的基础上，包括：10 K小麦害虫图像（Mutus-Tech），世界领先的害虫检测模型（PestNet; UoS），农艺和害虫管理专业知识（ADAS）。这种新的解决方案将提高农场的弹性，减少不必要的杀虫剂使用，并改善害虫控制。这项技术还将降低农民的种植成本。最终，它将提高农业生产力和利润，并刺激英国市场的增长。

Worldwide, crops are threatened by invertebrate pests which cause feeding damage and transmit plant viruses. High levels of infestation can cause up to 80% yield loss. While pesticides are often applied to crops protection, their uncontrolled application can cause soil erosion and contamination. Sustainable management of pest is reliant on: accurate identification of the pest present, knowledge of the levels of pest damage that can be tolerated, and effective pest management solutions for maintaining soil health. Currently, there is no integrated in-field solution for sustainable management of wheat pest in the UK.As restrictions on insecticide use increase and a greater number of insecticide resistant pest populations emerge, growers are looking towards more sustainable integrated pest management (IPM) practices. However, there are numerous barriers that restrict the uptake of IPM principles: Accurate identification of invertebrate pests is difficult and requires taxonomic training, a skill that growers often lack; current thresholds have received little testing and validation under field conditions, limiting grower confidence, etc. In an initial project we developed an early-stage solution to this problem by building an AI-driven mobile pest-detection solution to identify insect pests in wheat crops (Innovate project 10002902). The outcome could be further improved to be more successful with benefits below:* Expanding wheat pest detection model to above-group crops like potato and rapeseed.* Improving recognition accuracy and efficiency of models running in mobile phones with optimisation mechanisms.* Evaluating and validating the accepted pest thresholds for the focal crop pests.Here, we will propose an follow-on project that improves our AI-driven mobile pest management solution through optimising deep learning detection models with knowledge distillation technique to pests of other arable crops. The output will be an enhanced MPM solution that offers: 1) rapid detection and accurate quantification of foliar pests using mobile devices; 2) placing pest quantification into context of region-specific pest tolerance thresholds; 3) Providing estimation of economic thresholds and advice on pest control. It will build on existing resources in the consortium, including: 10K wheat pest images (Mutus-Tech), a world-leading pest detection model (PestNet; UoS), agronomic and pest management expertise (ADAS).This new solution will improve farm resilience, reduce unnecessary insecticide use, and improve pest control. The technology will also reduce growing costs for farmers. Ultimately, it will improve farm productivity and profits, and stimulate growth of the UK market.