應用卷積神經網路於田間茶葉病蟲害影像之分類

Abstract

茶葉病蟲害會對茶樹的生長造成危害，此不利的條件會導致茶樹枯萎進而影響茶的產量及利潤。若能在早期發現，農民即可做出適當的病蟲害管理，並控制病蟲害的災情來降低產量的損失。本研究在台灣中部及北部收集了1842張茶葉病蟲害之影像來建構影像資料庫。其中1429張影像被用來訓練一個能辨識不同病蟲害之更快速區域卷積神經網路(Faster-RCNN)模型，並選用VGG16為卷積模型。此模型能辨識赤葉枯病、茶餅病、藻斑病等三種病害，及斑潛蠅、薊馬、茶捲葉蛾、茶姬捲葉蛾、盲椿象等五種蟲害，並針對赤葉枯病和茶餅病進行危害面積比例的計算。在413張測試的影像當中，選用兩個評估參數，交集與聯集比 (Intersection over Union) 與信心值(Confidence Score)。當兩個參數分別被設定為0.5及0.05時，可以得出精確率(Precision)、召回率(Recall)、平均精確率(Mean Average Precision, mAP)為70.9%、80.0%、75.1%，並適合進行手動的病蟲害偵測。其中茶餅病和斑潛蠅的精確率高達了84.25%和94.35%。當信心值提升到0.5時，精確率、召回率、平均精確率變為83.1%、70.9%、66.9%，並適合運用在自動化的即時監測。在評估面積計算的部分，赤葉枯病、茶餅病、葉片的影像分割的準確率達到了88.46%、91.09%、88.03%。此一田間茶葉病蟲害判別模型的開發，可作為協助茶農進行即時監測、判讀病蟲害的發生狀態的便利輔助工具。

Tea (Camellia sinensis (L.) O. Kuntze) leaf lesions are detrimental to the growth of tea crops. The adverse events result in illness of tea leaves and causes direct reduction in yield and profit. Thereby, early detection or on-site monitoring of tea tree lesions can provide effective Integrated Pest Management (IPM) strategies to control the infected area and prevent further yield decreasing. In this study, 1842 lesion images were collected from northern and middle Taiwan to build the image database. From the database, 1429 images of tea leaves were used to train the model based on faster region-based convolutional neural network (Faster R-CNN) with VGG16 as the backbone model. The proposed model classifies three types of tea diseases: brown blight, blister blight, algal leaf spot and five types of insect pests: leaf miner, tea thrips, tea leaf roller, small tea tortrix, tea mosquito bug and calculate the area proportion of brown and blister blight. When the setting of two evaluation components, Intersection over Union (IoU) and confidence score, were set as 0.5 and 0.05, the results of 413 testing images obtained a precision, recall and mean average precision (mAP) of 70.9%, 80.0% and 75.1%, respectively. In this case it is more suitable for manual detection. In addition, the AP of blister blight and leaf miner reached up to 84.25% and 94.35%. When the confidence score was changed to 0.5, the precision, recall and mAP were 83.1%, 70.9% and 66.9%, respectively and could be applied for automatic real time detection. In the evaluation of area calculation, the accuracy for brown blight, blister blight and leaf segmentation were 88.46%, 91.09%, and 88.03%, respectively. The developed tea lesion classification model provides tea farmers a convenient tool for real time monitoring in the occurrence of tea field lesions.