以深度學習偵測植物病害之研究－以稻熱病為例

Abstract

稻熱病作為水稻之主要病害之一，目前防治稻熱病之主要方法為抗病品種之栽培以及殺菌劑之施用，雖然前者具有較低之生產成本，且對於環境所造成之負擔也較低，然而過度依賴單一品種或是不當的施肥管理，將使其抗病性逐年下降。而殺菌劑之施用雖然較為有效，則需要注意施藥之劑量以及範圍，避免過度施藥所增加之生產成本及環境衝擊。\n\n傳統在執行稻熱病害的偵測時多以人力進行地面調查，無論是時間上或是金錢上都需要花費大量成本，對於病徵判斷之精確度亦有限。隨著機器學習技術之發展，如深度學習這類具有高學習力及高辨識力之技術，可以提供使用者一次性進行大量資料的訓練，且其”end-to-end”之特性使資料不須進行預處理，可以節省大量時間成本，因此本研究欲使用深度學習之技術開發一套稻熱病病徵之影像辨識系統，先針對田野調查所拍攝之影像進行專家判釋，建立病害病徵之影像資料庫，並用其進行深度學習，建立稻熱病辨識之卷積神經網路模型。同時，本研究使用經紅外線改機之相機影像進行深度學習，探討此類資料對於稻熱病之判釋能力。\n\n本研究根據所蒐集之資料，將影像分為健康、稻熱病及胡麻葉枯病三個分類，並將可見光及紅外光影像分別進行深度學習之模型訓練。模型進行遷移學習以提升訓練效率，並使用DenseNet121作為模型架構。模型完成訓練後進行測試資料之預測，並以混淆矩陣進行辨識成果的展示，同時計算相關指標以進行成果精度評估。結果顯示利用可見光影像進行訓練之分類模型精度可達0.98，紅外線影像之分類模型精度達0.94，可見光影像在病徵辨識上較有優勢，而透過紅外線改機所獲得之紅外線影像亦能取得不錯的辨識精度，顯示此類影像應用於病徵辨識之潛力。

Rice blast is the major disease of rice in Taiwan, the current main method to control rice blast are the application of blast-resistant cultivars and use of fungicide. Although the former method has lower cost and lower impact to the environment, over-producing single cultivar or improper fertilizing management might cause the disease resistance decreasing. On the other hand, though the application of fungicide is more effective, it has concern of the dosing amount and area, to avoid the growing producing cost and environmental impact.\n\nTraditionally, diagnosis of rice blast is usually done by manual, which cost highly in either money or time, and the precision of the diagnosis is also limited. As the progression of machine learning, technique like deep learning that has high learning and recognizing ability can be supplied for training with large database, and its “end-to-end” characteristic can provide learning solution without data pre-processing which can highly speed up the procedure. As a result, this study used deep learning technique to develop a image recognition system for rice blast lesion, by manually interpreting the images from fieldwork through expert, and establish an image database of disease symptoms, then used it for deep learning training to develop a CNN model for rice blast recognition. Meanwhile, this study combined images from modified infrared-enabled camera for deep learning, the reliability of these data for rice blast recognition was tested.\n\nThe acquired data was divided into three classes including healthy, rice blast and brwonspot, and both visible light and near-infrared images were trained by a CNN model respectively. The training process was executed by transfer learning to reduce training cost, and a DensNet121 structure was implemented as the model structure. Examination of model performance was done with confusion matrix to display the recognition result, and related indices were calculated for futher estimation. The results show that the accuracy of visible light model is about 0.9812, and the accuracy of near-infrared model was about 0.9400. According to the result, the visible light image has better performance for rice blast lesion detection, and the near-infrared image obtained by modified camera has great potential for related utilization.