High-resolution UAV-based blueberry scorch virus mapping utilizing a deep vision transformer algorithm

• The first study to use UAV imagery for mapping blueberry scorch virus. • A new deep learning algorithm with superior classification performance. • The new algorithm illustrates superior spatial transferability. • This research helps to control the virus and thus generates high economic values. Blueberry scorch virus (BIScV), transmitted by aphids, causes a serious disease in highbush blueberries with a significant economic impact. Early detection and mapping of the distribution of BIScV infected plants in fields are critical to implementing effective disease management practices, such as the timely removal of infected bushes and control of aphid vectors. The conventional visual plant assessment for symptoms remains dominant in BIScV detections, though it is labor-intensive, time-consuming, and costly. In recent years, the use of remote sensing techniques has become popular for in-field assessments of crop diseases and insect pests incidence, and thus provides an effective approach for detecting and mapping BIScV infections. Convolutional Neural Networks (CNNs) are among the most widely employed algorithms in remote sensing image classification. However, CNNs have some limitations in their ability to obtain global information dependency due to the convolution’s constrained receptive field in each layer. To address this challenge, the self-attention mechanism utilized in Vision Transformers (ViTs) was suggested in previous studies for achieving flexible global information dependency through facilitating communication among arbitrary pixels in images. As such, we developed a CNN-ViT-based deep learning algorithm (named “Scorch Mapper”), a pixel-based classifier, that utilizes both the functionality and capabilities of CNNs in capturing local visual characteristics and ViTs for acquiring long-range information dependency for the mapping of BIScV. We also compared the developed Scorch Mapper to several other CNN– and ViT-based algorithms, including a 2D CNN, ResNet, HybridSN, Swin Transformer, Efficient Net, CMT, InFormer, and Efficient Former. Our results demonstrated the superiority of the Scorch Mapper compared to other CNN– and ViT-based algorithms. Research findings also show that the Scorch Mapper is effective and can be applied over a wide area to support BIScV mapping and monitoring. Furthermore, the developed model opens a new window for future automatic BIScV mapping utilizing cutting-edge remote sensing algorithms and technologies.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.