建立植物病蟲害影像辨識模型-以葡萄為例

Abstract

葡萄是一種具有經濟價值的農作物，但其生產和品質常受到病害與蟲害的嚴重影響，如銹病、露菌病、白粉病、葉螨和薊馬等。這些疾病不僅減少了產量，還降低了葡萄的品質，進而影響到整個供應鏈的經濟效益和穩定性。傳統的病害檢測方法依賴大量的人力和時間，這限制了其效率和應用範圍。 為提高檢測的效率和精度，引入影像辨識技術成為了一種有效的解決方案，這不僅能節省成本，還提高了檢測的自動化程度。基於深度學習的影像辨識技術在目前被廣泛應用於植物病害的識別工作中。然而，目前的研究多依賴於實驗室條件下的影像數據進行模型訓練，這限制了模型在實際田間環境中的適用性。此外，少數基於田間條件的研究雖然使用大量實際數據進行訓練，但這一過程不僅耗時，也需要較高的成本。這些挑戰顯示，將深度學習技術實際應用於植物病害檢測仍面臨諸多困難。 為了解決實際層面的應用問題，本研究旨在探討如何以稀疏的田野條件資料集建立出具有優秀泛化能力的影像辨識模型。本研究結合物件偵測模型與分類模型的優勢，構建一個新穎的辨識框架，藉以更高效率的完成分類目標。本研究評估了多種近期表現突出的物件偵測演算法，以選擇最適合本研究資料集的模型。對於框架中的分類模型，本研究提出了一個新穎的分類模型Vit- LRK，該模型結合卷積結構與變換器結構的優勢，且當中運用了遷移學習和模型集成的技術來提升表現。同時，本研究也提出了一個嶄新的訓練策略，旨在進一步提升模型的泛化能力。模型效能的評估則是使用了佔整體資料集33%的測試集，以確保結果的統計可靠性。 結果顯示，Dino 在測試資料集上的表現優於本研究評估的其他先進物件偵測模型。當進一步將Dino與不同的分類模型進行整合並評估其性能後， Dino與Vit-LRK形成的混合模型，在性能表現上不僅明顯超越了單獨運用 Dino 的效果，同時也顯著優於將Dino與從植物疾病檢測相關文獻中篩選出的10個現有分類模型整合的結果。 綜合以上論述，本研究開發了一個新穎的影像辨識模型，能夠在資料稀疏的條件下，透過使用實地拍攝的葡萄疾病影像進行訓練。本模型不僅適應了有限的數據集，還展現出了卓越的泛化能力。此成果證明了本研究的方法在處理現實世界中的農業疾病辨識問題上具有重要的應用潛力。

Grapes are a valuable agricultural crop but their production and quality are often severely affected by diseases and pests such as rust, downy mildew, powdery mildew, spider mites, and thrips. These afflictions not only reduce yield but also degrade the quality of the grapes, impacting the economic benefits and stability of the entire supply chain. Traditional methods of disease detection rely heavily on manual labor and time, limiting their efficiency and scope. To enhance the efficiency and accuracy of detection, the introduction of image recognition technology has proven to be an effective solution. This not only saves costs but also increases the level of automation in detection. Currently, image recognition technologies based on deep learning are widely applied in plant disease identification. However, most studies rely on images under laboratory conditions for model training, which limits the applicability of the models in actual field environments. Additionally, although a few studies based on field conditions use a large amount of real data for training, this process is not only time-consuming but also costly. These challenges show that the practical application of deep learning technologies in plant disease detection still faces many difficulties.

To address the practical application issues, this study aims to explore how to build an image recognition model with excellent generalization abilities using sparse datasets from field conditions. This study combines the advantages of object detection and classification models to construct a novel recognition framework that efficiently achieves classification objectives. The study evaluated various recent object detection algorithms to select the most suitable model for our dataset. For the classification model within the framework, a novel model called Vit-LRK was proposed, combining the advantages of convolutional and transformer architectures and employing transfer learning and model ensemble techniques to enhance performance. A new training strategy was also proposed to further improve model generalization. The effectiveness of the model was evaluated using a test set that accounts for 33% of the total dataset to ensure the statistical reliability of the results. The results show that Dino performed better on the test dataset than other advanced object detection models assessed in this study. When Dino was further integrated with various classification models and their performance evaluated, the hybrid model formed by Dino and Vit-LRK not only significantly outperformed using Dino alone but also surpassed integrating Dino with ten existing classification models selected from the literature on plant disease detection.

In summary, this study developed a novel image recognition model that, trained on real-world images of grape diseases under sparse data conditions, not only adapts to limited datasets but also demonstrates excellent generalization capabilities. These results prove the significant potential of our approach for addressing real-world agricultural disease identification problems.