請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/8538
標題: | 應用無線影像環境感測網路於智慧整合蟲害管理 Intelligent and Integrated Pest Management Using Wireless Imaging and Environmental Sensor Networks |
作者: | Dan Jeric Arcega Rustia 羅傑瑞 |
指導教授: | 林達德(Ta-Te Lin) |
關鍵字: | 整合蟲害管理,黏蟲紙,影像辨識,深度學習,無線感測網路,數據分析, integrated pest management,sticky paper trap,image recognition,deep learning,wireless sensor network,data analytics, |
出版年 : | 2021 |
學位: | 博士 |
摘要: | 為了推動農民使採用智慧化的蟲害整合管理(IPM),本研究開發了一套可以自動監測農業生產場域中害蟲數量及環境相關資訊的智慧整合系統。本系統由無線感測器裝置所組成,這些裝置會將黏蟲紙的影像傳送至遠端伺服器進行分析並回傳訊息給農民作為蟲害整合管理之應用。本研究同時開發了用於黏蟲紙影像上自動偵測及辨識害蟲之演算法,該演算法利用卷積神經網路(CNN)深度學習模型以級聯的方式進行害蟲偵測和分類;分別是由一個負責從黏蟲紙影像上定位出物件的物件偵測器模型及一個用於辨識所偵測到的物件種類的分類模型所組成。深度學習模型根據安裝場域的特性分別以兩種不同的深度學習方法來進行訓練。對於溫室場域應用了半監督式學習的技術,利用系統新擷取的黏蟲紙影像自動收集害蟲訓練影像並重新訓練害蟲分類模型。經過一年的持續訓練,分類模型基於物件層次和影像層次的測試F1分數都可達到0.93。至於戶外場域則是對多個分類模型進行訓練並根據分類學進行級聯,將害蟲影像分類到物種級別,此級聯演算法基於物件層次和影像層次測試的平均F1分數分別為0.91和0.89。本系統已安裝在多個農業生產場域中驗證與測試並提供農場管理者進行評估與應用於蟲害整合管理。本研究並透過數據分析從長期資料中獲取相關資訊開發了警報模型及生物學模型,將蟲害數量數據整理轉換為有意義的資訊,例如可以提供蟲害整合管理行動建議的警報級別,以幫助農場管理者進行決策。在對所有實驗場域的測試中,利用所訓練的生物學模型來描述害蟲的飛行頻率行為,平均r2值可以達到0.97。這些整理過的資訊是通過網站和行動APP與農場管理者共享以便有效應對蟲害狀況。根據農場管理者的使用情況及回饋意見顯示,本系統能夠透過量化的參考依據來協助擬定蟲害整合管理策略,而自動化系統能夠讓研究人員及專家引導農民進入數據驅動和友善環境的蟲害管理。 To drive farmers into smarter integrated pest management (IPM), an integrated and intelligent system that can automatically monitor the number of insect pests and measure environmental conditions in agricultural production sites was developed. The proposed system is composed of wireless sensor nodes that send images of sticky paper traps to a remote server. An algorithm was developed to automatically detect and recognize insect pests from the sticky paper trap images. The algorithm features a cascaded approach for detection and classification using convolutional neural network (CNN) deep learning models. It is composed of an object detector model, which locates the objects from the sticky paper trap images, and image classifier models that identify the detected objects. The deep learning models were trained using two different deep learning methods fitted for each type of installation site. For indoor sites such as greenhouses, a semi-supervised learning technique was applied to train a multi-class insect classifier model. The proposed technique was used to automatically collect training images from newly acquired sticky paper trap images and retrain the classifier model. After a year of continuously training the classifier model, F1-scores of 0.93 was achieved on testing both by object level and image level. For outdoor sites like orchards, multiple classifier models were trained and cascaded taxonomically to classify insect pest images up to the species level. The cascaded algorithm was found to have average F1-scores of 0.91 and 0.89 by object level and image level testing, respectively. The system was installed in several agricultural production sites for farm managers to evaluate and utilize in their IPM routines. Data analytics was applied to extract information from the acquired long term data. Alarm models and biological models were developed that convert insect pest count data into valuable information such as alarm levels that indicate action recommendations for IPM, guiding farm managers in decision-making. Upon validation, the biological models were able to describe the flight rate behavior of the insect pests with an average r2 of 0.97 based on the fitted models of all the experimental sites. The information was shared through a website and mobile APP which farm managers used to effectively respond to the present insect pest condition. The usage information and feedback given by the farm managers showed that the system was able to help them by having quantitative reference for their IPM strategies. This research presents different ways an automated system in assisting farmers for IPM applications, which can be used by researchers and experts in bringing farmers to data-driven and environmentally friendly insect pest management. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/8538 |
DOI: | 10.6342/NTU202100237 |
全文授權: | 同意授權(全球公開) |
電子全文公開日期: | 2024-01-01 |
顯示於系所單位: | 生物機電工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2801202118555100.pdf | 15.66 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。