應用自動化監測系統於非疫生產點之可行性研究

Abstract

瓜實蠅及東方果實蠅等實蠅類害蟲為歐、美、日、韓等國家的重要蔬果檢疫害蟲，農產品進出口皆必須面臨繁複的檢疫規定。為了克服這些檢疫障礙，確保農產品的順利流通，其中的一個方法即是在實蠅類害蟲疫區建立非疫生產點，以確認此區域沒有特定害蟲存在的系統性管理策略，來監測害蟲並維持其非疫狀態。 在非疫生產點進行人工害蟲監測工作，若無法即時對田間害蟲發生狀況進行回報，將對緊急防治技術操作形成一資訊落差。本研究藉由物聯網害蟲監測系統實地進行區域性的害蟲密度監測。主要使用四套設備，其一為害蟲誘引器搭配自動化害蟲計數蟲桶，以監測非疫生產點溫室外的害蟲密度，及監測溫室內的蟲害入侵情形。其二為害蟲影像拍攝裝置，定時對黃色黏蟲紙進行拍攝，並回傳黏蟲紙影像至雲端資料庫，搭配害蟲影像辨識模型，監控非疫生產點的戶外害蟲種類及數量，及監測溫室的雙重門間是否有害蟲入侵。其三為自動化害蟲匯報系統，可每日發送自動計數蟲桶及害蟲影像拍攝裝置的資訊，統計每日內各監測點所累計的蟲數，及發送溫室內害蟲入侵的警報。其四為場域安全性智慧通報系統，自動化記錄人員進入溫室的次數及在溫室內進行的農事作業，人員進出雙重門動態及大型農機具於鐵捲門進出時的畫面也會被紀錄並上傳至雲端資料庫。 自動害蟲計數蟲桶能即時將進入監測陷阱之果實蠅類害蟲上傳雲端資料庫，此設備應用於場域時期達12個月，資料回傳率均超過95%。害蟲影像拍攝裝置於戶外監測點及雙重門間拍攝黏蟲紙，並回傳害蟲動態資訊，在雲林、彰化場域，設備回傳率超過96%。害蟲影像辨識模型使用雙向特徵金字塔及多尺度注意力網路結合YOLOv8n，並使用遷移式學習訓練得到最好的訓練結果，以臺大農場收集之黏蟲紙影像進行辨識，得到瓜實蠅精確率0.930，召回率0.943；東方果實蠅精確率0.896，召回率0.908，瓜、果實蠅總共的漏抓率僅0.027。以非疫生產點實際的黏蟲紙影像進行辨識，準確率達90.1%。 在蟲害防治成效方面，嘉義場域連續12個月未發現瓜、果實蠅，達到非疫生產點要求，而雲林場域曾在某一次調查週期內出現2隻雄果實蠅入侵，經屋頂的修繕和噴藥處理後，即不再有入侵的紀錄，可繼續維持非疫生產。此外，系統在技術、經濟和操作方面均具有可行性，使用自動化監測成本低於傳統的人工監測成本，也為此自動化系統制定一套標準作業程序供農民及研究人員使用。本研究成功建立了一套自動化害蟲監測系統，為非疫生產點的建立及維持提供了有力支撐，希望藉由非疫生產模式將小果番茄等無法以蒸熱或低溫檢疫處理的農產品，出口至國外。

Melon fruit flies and oriental fruit flies are important quarantine pests for countries such as Europe, America, Japan, and South Korea. Agricultural products must face complex quarantine regulations for import and export. In order to overcome these quarantine barriers and ensure the circulation of agricultural products, one of the methods is to establish pest-free production sites in areas affected by fruit flies. This involves a systematic management strategy to confirm that there are no specific pests present in the area, monitoring the pest, and maintain its pest-free status. It is difficult to promply report on pest presence if pests are monitored manually in pest-free production sites, and this will create an information gap in pest control. This study conducts pest monitoring and detection in pest-free production sites through an Internet of Things (IoT)-based pest monitoring system. It mainly uses four sets of equipments. The first is a pest attractant paired with an automated pest counting device to monitor the density of pests outside the greenhouse in pest-free production sites, and to detect pest invasion inside the greenhouse. The second is a pest image capture device that takes photos of yellow sticky traps in a fixed intervals and sends the images to a cloud database. Then, a pest image recognition model established for identifying the types and quantities of outdoor tephritid near pest-free production sites, and to monitor whether pests have invaded between the double doors of the greenhouse. The third is an automated pest reporting system that can send out the data collected by the automated pest counting trap and pest image capture device daily. The system accumulates the number of pests at each monitoring point from the previous day, and sends alerts for pest invasions inside the greenhouse to users. The fourth is a smart aleart system that automatically records the number of personnel entering the greenhouse and the types of agricultural tasks conducted in the production site. The system also records and uploads the events of personnel entering and exiting the double doors and large agricultural machinery entering and exiting through the rolling doors to a cloud database. The automatic pest counting trap can instantly upload the data of fruit flies entering the monitoring traps to the cloud database. This device is used in field for 12 months, with a data return rate exceeding 95%. The pest image capture device captures images of sticky traps placed outdoors and between double doors, while transmitting dynamic pest information. In the Yunlin and Changhua experimental sites, the device has a data return rate of over 96%. The pest image recognition model uses bidirectional feature pyramids and efficient multi-scale attention networks combined with YOLOv8n, and achieves the best training results through transfer learning. Trained by the images of sticky traps collected at the National Taiwan University farm, the proposed model achieves a precision rate of 0.930 for melon fruit flies and a recall rate of 0.943. For oriental fruit flies, the precision rate is 0.896, with a recall rate of 0.908. The total miss rate for melon and fruit flies is only 0.027. When recognizing the images of sticky traps from pest-free production sites, the accuracy rate reaches 90.1%. In terms of pest control effectiveness, the Chiayi experiment site has not found any melon or oriental fruit flies for 12 consecutive months, and this meets the requirements of a pest-free production site. During a survey period, the Yunlin site found two invaded male fruit flies, but after roof repairs and pesticide treatment, there was no longer be any records of invasions, and it was able to continue maintaining pest-free production site. Furthermore, the system is feasible in terms of technology, economics, and operations. The cost of using the automated monitoring system is lower than traditional manual monitoring costs. A set of standard operating procedures has been established for farmers and researchers to use this automated pest monitoring system. This study has successfully established an automated pest monitoring system and provided strong support for the establishment and maintenance of pest-free production sites. It is hoped that through the pest-free production sites, agricultural products such as cherry tomatoes that cannot be treated with steam heat or low-temperature quarantine can be exported abroad.