AIoT技術在設施栽培中精密監控小型害蟲與優化蜜蜂授粉之應用

Abstract

在臺灣，為提高作物單位面積產量，通常採用易於管理的溫室種植方法。然而即便作物採用溫室種植，仍無法完全避免蟲害爆發。另一方面，蜜蜂不太適應於密閉空間內飛行，進而導致作物授粉效率不太穩定。因此，在蟲害防治方面，基於過去研究，本研究利用物聯網系統搭配Faster R-CNN、YOLO系列以及Transformer系列物件偵測模型進行大數據分析，我們提出更加輕量化且客製化的方法，以提供溫室害蟲防治之有效參考指標；另外授粉效率評估方面，過去研究多利用蜜蜂進出巢數量評估作物授粉效率，此法無法得知蜜蜂攜帶花粉粒大小，進而無法精確掌握與授粉效率節節相關的花粉量，因此本研究提出利用Mask R-CNN、YOLACT以及Transformer系列等實例分割模型的方法，針對花粉粒面積進行辨識以及計算，並進一步與實際花粉粒重量進行相關性分析，以優化蜜蜂授粉。基於以上研究成果，將對溫室中種植作物之蟲害防治以及蜜蜂授粉優化提供一個較智能的解決方案，進而提昇作物產量。

In Taiwan, to increase the yield per unit area of cultivation, greenhouse cultivation methods that are easy to manage are usually adopted. However, even if crops are grown in greenhouses, pest outbreaks cannot be completely avoided, and bees are not well adapted to flying in confined space, resulting in unstable pollination efficiency of crops. Therefore, in terms of pest control, based on past research, this research uses an IoT system with Faster R-CNN, YOLO series, and Transformer series object detection models to conduct big data analysis. We propose a more lightweight and customized method to provide greenhouse pest control with effective reference indicators. In addition, in terms of pollination efficiency assessment, past studies mostly used the number of honey bees entering and leaving the hive to evaluate crop pollination efficiency. Such methods cannot know the size of pollen grains carried by honey bees and thus cannot accurately grasp the amount of pollen-related to pollination efficiency. The study proposes a method of instance segmentation models such as Mask R-CNN, YOLACT, and Transformer series to detect and calculate the pollen grain area and conduct correlation analysis with crop pollination to provide a more reliable pollination efficiency evaluation index. Based on the above research results, a more intelligent solution will be provided for pest control and pollination efficiency assessment of crops grown in greenhouses, thereby maximizing crop yields.