An attention-based deep learning model for early detection of polyphagous shot hole borer infestations in plants.

Abstract

The Polyphagous Shot Hole Borer (PSHB) is a highly invasive beetle that has been spreading like an epidemic across agricultural and forestry landscapes in recent years. Its rapid and destructive spread has turned it into a major global threat, causing widespread damage that continues to grow with time. Countries like South Africa, the United States, and Australia have implemented extensive measures to control the spread of PSHB, including the establishment of specialized agricultural support centers for early detection. However, there is still a strong need to make PSHB detection more accessible, allowing even non-experts to easily identify infections at an early stage. Artificial Intelligence (AI) has shown great promise in plant disease detection, but a major challenge in the case of PSHB was the lack of a suitable dataset for training AI models. In the proposed work, we first created a dedicated dataset by collecting images of trees infected with PSHB. We applied a range of preprocessing techniques to refine the dataset and prepare it for AI applications. Building on this, we developed a novel AI-based method, where we trained a deep learning model using a multi-convolutional layer network combined with a Fourier transformation layer. Additionally, an attention mechanism and advanced feature extraction techniques were incorporated to further boost model performance. As a result, the proposed approach achieved an impressive top accuracy of 92.3% in detecting PSHB infections, showing the potential of AI to offer a simple, efficient, and highly accurate solution for early disease detection.

Plant phenotyping relevance

植物画像から感染状態を推定するデータセットと深層学習手法を開発し、性能評価まで行っており、病害フェノタイピング手法が中心である。

Code and data availability

The paper's PSHB image dataset is not publicly available; the data availability statement says it can only be requested from the corresponding authors. No public code, model checkpoints, or dataset URLs are provided.

No evidence-backed public reproduction asset is currently recorded.