應用多個深度學習模型和智慧型手機自動辨識番茄病害、蟲害和生理障礙

Abstract

番茄是蔬果兩用的經濟作物。然而番茄產量與品質受到來自生物性的病蟲害與非生物性的生理障礙的影響，正確植物病理診斷為田間施藥與經營管理的第一步。目前，慣行農法的植物病理診斷多仰賴農友或植物病理專家的經驗與專業，難以進行即時且高通量的病蟲害與生理障礙辨識。因此，一套自動辨識病蟲害與生理障礙辨識的工具，將有望改善台灣番茄生產面臨的困境。本研究旨在開發一套番茄病蟲害與生理障礙辨識模型，可透過手機拍攝番茄葉片影像辨識台灣番茄生產常見的病蟲害與生理障礙，並提供即時的防治處方。本研究至田間拍攝共12,673張受病蟲害與生理障礙侵擾之番茄葉片影像用於訓練深度學習模型。侵擾之病蟲害與生理障礙皆經過植物病理學家、應用動物學家和園藝學家鑑定。此系統之辨識服務由五個深度學習模型構成階段式辨識策略，其透過「異常影像偵測模型」用於辨識輸入之影像是否為番茄葉片，「葉片分類模型」用於區分羽狀複葉和小葉，「羽狀複葉診斷模型」用於辨識羽狀複葉是否受病毒感染或是生長激素噴灑傷害，「小葉診斷模型」用於辨別番茄葉片上病徵的位置和病蟲害與生理障礙類別，「葉黴病與新型白粉病分類模型」用於區分葉黴病與新型白粉病。經過訓練之五個深度學習模型準確率分別達93.00%、95.05%、99.11%、96.22%和98.07%。本研究開發一個聊天機器人與辨識服務整合，使用者可以透過手機自動辨識番茄病蟲害與生理障礙。本研究提出一個提供自動辨識田間番茄病蟲害與生理障礙之方法，使番茄病蟲害與生理障礙之辨識更便利，協助農民迅速採取即時的應對措施。

As an essential food in the culinary cultures of many countries, tomato is one of the commonly cultivated crops worldwide. The yield and quality of tomato, however, can be significantly impacted by diseases, pests, and disorders (DPD). Most DPD symptoms (e.g., spots, yellowing, necrosis, and leaf distortion) exhibit on leaves first. The symptoms can be confounded and confusing at a certain level. Thus, correctly identifying the cause of a symptom is crucial for tomato cultivation management. Conventionally, the cause of a symptom was identified using naked-eye or laboratory-based examination by experienced farmers or experts. These approaches are time-consuming and highly rely on expertise. However, immediate actions may need to be taken in crop management. Thus, this study aims to provide a rapid and automatic solution for identifying tomato DPD symptoms exhibited on leaves using deep learning models and smartphones. In this research, 12,673 images of tomato leaves exhibiting symptoms of DPD were gathered under field conditions using digital cameras and smartphones. The causes of the symptoms in images were curated by plant pathologists, entomologists, and agronomists. Five deep learning models were trained to identify tomato DPD automatically, including: (1) anomaly detection model (ADM) for verifying if received images contain tomato leaves, (2) leaf type classification model (TCM) for differentiating between pinnate compound (PC) leaves and leaflets, (3) compound leaf diagnosis model (CDM) for identifying virus and phytohormone damage on PC leaves, (4) leaflet diagnosis model (LDM) for identifying 13 categories of DPD on leaflets, and (5) leaf-mold-or-powdery-mildew-II model (LPM) for determining if yellow spots are caused by leaf mold or powdery mildew II. A chatbot controller was developed to facilitate communication between users and models through smartphones, making the identification fully automatic. ADM, TCM, CDM, LDM, and LPM achieved an accuracy of 93.00%, 95.05%, 99.11%, 96.22%, 98.07%, respectively. The proposed approach simulated the logic of experts in tomato issue diagnosis and can assist in tomato cultivation management in the field rapidly.