應用深度學習演算法於胡瓜葉表複合病害及病程辨識系統之開發

Abstract

胡瓜為世界上高經濟價值的作物之一。病蟲害為造成其產量損失的主因之一。在遭病蟲害感染初期，若能正確識別危害源，則能儘速採取應對措施。罹病植株葉表多會呈現病斑徵狀，其病斑依其病蟲害種類、病程有所差異及變化。初期表徵通常為較不明顯的病斑，易與健康葉片或其他類輕微病害混淆。另，亦有可能同時感染多種病害，且因而產生更為複雜的病斑表徵。病蟲害類別判讀通常由植病專家進行，然所需專業門檻高，專業人才稀缺。本研究目標為應用深度學習方法開發一套自動判別病蟲害類別、病程，及複合病害之判讀系統，並藉由串聯聊天機器人功能提供使用者相關服務。本研究的判別類別涵蓋健康葉片、七種單一病害及十種複合病害，病程方面則分為早、中、晚三期。影像資料集均為田間實地拍攝之植株病蟲害影像，共計8000餘張。深度學習方法選用更快速區域卷積神經網路(Faster region-based convolutional neural network，Faster R-CNN)為主要架構。共建構兩種模型，一為單純使用Faster R-CNN進行分類預測的一步驟辨識模型，及結合Faster R-CNN與病程分類器的二步驟模型。於影像標記方面，採取one-hot及multi-hot兩種方法，比較其於複合病害的辨識效果，及對涵蓋類別進行擴增的應用彈性。模型優化方面，使用資料重採樣(resampling)、增益(augmentation)及骨架替換(backbone substitution)等方法。於最終模型開發結果呈現上，一步驟模型搭配multi-hot標記方法，可達0.846的F1-score, 0.758的平均精確度均值(Mean average precision，mAP)及0.733的準確率(accuracy)。於病程辨識上，一步驟模型可達0.648的F1-score，而二步驟模型則可達0.701的F1-score。以Gradient-based class activation map (Grad-CAM)與主成分分析(Principal component analysis，PCA)將模型中特徵表現進行可視化分析，均顯示於病程預測上，二步驟模型的病程分類器可強化類別特徵。此胡瓜病蟲害識別模型並串接至聊天機器人應用服務端，提供便利使用識別功能及回饋意見的使用者介面。藉由此一系統的開發，將有望協助農場管理人員即時判讀植株健康狀態，降低由病蟲害所造成之產量損失。

Cucumber (Cucumis Sativus L.) is one of the most important crops in the world. Cucumber diseases are one of the causes of annual production and yield losses. If the disease can be correctly identified in the early stage, then measures can be taken to eliminate it timeously. Lesion patterns typically occur on the foliar surface of cucumber leaves, and they may vary based on the type and progress of the disease. In early stages, diseases usually cause relatively unclear foliar patterns that are easily confused with healthy leaves or other minor early diseases. Moreover, cucumbers can be infected with multiple diseases and show complicated patterns simultaneously. Diseases are traditionally identified by professionals, and the services are in high demand; however, such people are rare. Thus, this study aims to develop an automatic identification system for identifying the disease type, progress, and multi-disease cases, and connect an instant-message bot service for practical use. In this study, there were healthy leaves, seven single diseases, and ten multi-diseases. The disease progress was categorized into the early, middle, and late stages. Approximately 8,000 field images were collected by cameras and smartphones. The selected deep learning neural network was the faster region-based convolutional neural network (Faster R-CNN). Two models were constructed, a one-step model that is a single Faster R-CNN, and a two-step model composed of Faster R-CNN and a disease progress classifier. For image annotation, two labeling methods, one-hot and multi-hot labeling, were applied and compared on the prediction performance of multi-disease cases and flexibility of expansion on disease types. For model optimization, data resampling, augmentation, and backbone substitution were implemented. The final results showed that the one-step model with multi-hot labeling achieved an F1-score of 0.846, a mean average precision (mAP) of 0.758, and an accuracy of 0.733. For the prediction performance of disease progress, the one-step model obtained an F1-score of 0.648, whereas the two-step one achieved a better F1-score of 0.701. The gradient-based class activation map (Grad-CAM) and principal component analysis (PCA) were adopted to visually analyze the disease progress classifier. Both techniques showed that the disease progress classifier learned better features of disease progress. The identification system contained an identification model and an instant-message bot, and it provided the convenient functionalities including disease identification and opinion feedback. The development of this system is expected to help farm managers to timely acquire the health status of a plant and reduce the yield loss caused by diseases.