自動化太陽能害蟲監測系統之研究

字體大小:A- A A+

單位商業會員附件下載:
論文名稱: 自動化太陽能害蟲監測系統之研究
研究生姓名: 李維哲
指導教授姓名: 林達德
出版年: 2021
學校名稱: 國立臺灣大學
系所名稱: 生物機電工程學系
關鍵字: 太陽能系統;邊緣運算;深度學習;害蟲監測系統;Solar powered;Edge Computing;Deep learning;Insect pest monitoring system
摘要: 隨著環保意識抬頭,兼顧公眾健康、保護環境及有益生物的整合式病蟲害管理IPM (integrated pest management)逐漸為農民採用。IPM利用多元防治手法控制害蟲族群,其經濟、有效且永續的管理方式在為作物增添附加價值同時,也為大自然的永續發展盡一份綿薄之力。建立IPM的關鍵之一是對作物做全面性的監測,不只能即時掌握場域的害蟲發生情形,更可以透過長時間的追蹤害蟲數量、環境資訊,從中爬梳各因子間的因果關係來建立害蟲的生長模型、作物生產的經濟模型、製作害蟲爆發的預測預警系統。 以輔助IPM做為研究發想,本研究旨在建立太陽能供電的害蟲影像監測系統,系統藉由深度學習辨識黏蟲紙影像上的害蟲種類及數目,一併收集環境溫、溼、照度資料回傳至伺服器,將資訊以可視化數據呈現於網頁和手機app。將裝置部署於果園或溫室中搜集黏蟲紙和環境資訊可幫助農民全面了解該場域現況,並作為輔助蟲害防治決策使用。裝置由Arduino Pro mini和Raspberry Pi 4開發板組成,Pi camera v2拍攝黏蟲紙影像,影像依序經過YOLOv3-tiny做潛在害蟲偵測和 MobileNetv2做害蟲種類辨識,另紀錄太陽能和系統發用電量情形,討論該監測裝置實際功耗及實驗場域中太陽能發電的可行性和日照情形。所有資訊由Raspberry Pi 4透過無線網路傳回至伺服器儲存,透過自動計數演算法修正被誤判的害蟲種類,以累計圖呈現實驗場域的害蟲發生情況。 實驗場域共有台南新化農業改良場芒果園和高雄鳳山熱帶園藝試驗所芭樂園兩處,選定在台灣危害果樹最嚴重的果蠅科害蟲為標的害蟲,系統使用的潛在害蟲偵測模型YOLOv3-tiny的mAP@.5達93.56%,果蠅科害蟲辨識模型MobileNetv2的F1-score達0.94。系統用電情形為每天15W,可承受4天連續下雨等無日照情況。另外系統設計成方便組裝拆卸,得以讓使用者快速部署至場域,系統穩定性也經過實地驗證的考驗,裝置已和廠商合作技術轉移。There has been an great awareness about environmental protection in recent years, and the idea of integrated pest management (IPM) has gradually approved and implied by food producers. By combining multiple pest control strategies, IPM aims to build up a harmony management not only do good to species, do health to consumers but also make it a sustainable development style of plant production. One of key steps in IPM is to collect information about plants comprehensively. By learning more to the growth condition to plants, climate changes, circumstances of pest and disease, the more explicit model such as economic growth model or insect pest early warning model could be build from them. Inspired by the idea of IPM, this research aims to build a solar powered insect pest monitoring system. System collects environmental factors with Arduino and analyses the amount of insect pest on sticky traps through deep learning methods in Raspberry Pi 4, result would be sent to servers and be visualized-displayed on website or apps so that the users could take fully control the situation in fields in time easily. Edge computing composed of two stages, finding insect pest on sticky trap images using YOLOv3-tiny and classifying the species using MobileNetv2. Two experiments were being held in Tainan Xinhua DAREs mango farm and Kaohsiung Fengshan Tropical Horticultural Experiment branch guava field to study the stability . Tephritidae, also known as fruit fly, which do the most severe damage to orchard in Taiwan is selected as the prime target insect pest. The mAP@.5 of insect pest detection model is 93.5%, F1-score in fruit fly recognition is 0.94. Power usage for whole system is 15W per day and system can endure 4 contentious days without sufficient sunrise condition by designed and experiment result. Apart from these features, system is delicately designed to be easily assembled that allowed users to build up quickly in the field. The idea of the system has been technology transfer to the cooperated company.
資源連結: 前往查看