Precision identification of arecanut yellowing disease infected plants using UAV remote sensing and deep learning

Areca nut yellowing disease is one of the systemic diseases that severely threaten the growth of Areca catechu palms. It is crucial to precisely monitor the infected plants in the sustainable areca nut industry. Conventional monitoring has relied primarily on visual inspection and manual surveys. It is often required for the efficient operation under the frequent cloudy and rainy weather in Hainan Province, China. Unmanned Aerial Vehicle (UAV) remote sensing can be expected to serve as an effective solution for the high quality and availability of satellite optical imagery, due to the high spatial resolution, flexible response, and low costs. However, several limitations are still suffered from the monitoring of areca yellowing diseases, including the low identification accuracy and significant interference from similar ground objects, such as the discolored shrubs. This study aims to identify the areca nut yellowing disease-infected plants using UAV remote sensing and deep learning. Field sampling was also performed to assess the disease severity of Areca catechu palms, according to the disease features and the separability in the remote sensing imagery. Subsequently, a MicaSense RedEdge-M multispectral camera on a DJI Phantom 4 Pro V2.0 UAV was employed to acquire 5-band multispectral imagery, including the blue, green, red, near-infrared, and red-edge bands. Euclidean and Jeffreys-Matusita (J-M) distances were calculated to extract the sensitive spectral indices with high separability. Four datasets were constructed to train the typical models. A standardized dataset was then established from a total of 18,520 high-quality samples after image preprocessing, sample slicing, and manual verification. This final dataset was partitioned into training, validation, and test sets at a ratio of 8:1:1. A series of target improvements were implemented using the YOLO (You Only Look Once) v10 algorithm. A precise model was developed to identify the infected plants. Firstly, the lightweight GhostNet model was incorporated into the backbone network. The feature generation was enhanced to represent the effective information from the key multispectral bands. A redundant feature was reduced for the fine-grained disease textures. Secondly, a bidirectional feature pyramid network (BiFPN) was employed to replace the parts of the original structure at the feature fusion stage. Its learnable weights and bidirectional cross-scale fusion strengthened the multi-scale feature aggregation after modification, particularly in the complex forest stands. Finally, the Shape-aware intersection over union (SIoU) Loss function was introduced into the regression branch of the detection head. A more stable and geometrically constrained bounding box regression improved the localization accuracy of infected plants to accelerate the convergence of the model. The results demonstrate that the enhanced cross-scale feature fusion of the improved YOLOv10 model increased the mAP@0.5 to 90.6%, which was improved by approximately 20%, compared with the common models, such as Faster R-CNN, YOLOv8, and YOLOv10. Furthermore, the better performance was achieved in a precision of 91.2% and a recall of 92.7%, both of which significantly outperformed the classic deep learning models. The rapid and accurate detection of areca catechu plants infected with yellowing disease can be expected to effectively meet the practical demands for the inspection and monitoring of the disease.