Nondestructive detection of sweet potato leaf curl virus using 3D laser imaging combined with deep learning

Abstract

• Leaf curl virus detection in sweetpotato using molecular methods is costly. • Laser 3D scans of sweetpotato plants were used to test this concept. • PointNet++ used 3D point clouds for swift healthy vs diseased classification. • Data augmentation and downsampling enhanced model performance. • Results proved that this method can detect leaf curl virus rapidly and accurately. Sweet potato leaf curl virus (SPLCV) is a systemic viral disease of sweetpotato plants causing significant yield losses and posing a major threat to sweetpotato production. To achieve effective disease management, prompt detection of SPLCV is necessary and crucial. Currently, SPLCV is detected using a variety of molecular techniques, including polymerase chain reaction, loop-mediated isothermal amplification, and enzyme-linked immunosorbent assay, which are laborious, expensive, time-intensive, and limited in practicality for use in rapid, in situ diagnostic applications. Therefore, there is a need to establish alternative, point-of-care techniques for SPLCV detection. The goal of this study was to investigate the potential of using 3D point cloud data and deep learning to detect SPLCV. In this study, 3D point cloud data derived from multispectral laser scanning of healthy and SPLCV-infected sweetpotato plants were modeled with the PointNet++ neural network to discriminate healthy versus diseased plants. Spatial coordinates (X, Y, Z) and color information (R, G, B) of sweetpotato leaf point cloud data were used to develop and improve the efficiency of the classification process. The collected 3D point cloud dataset was first enhanced using the filtering approach to reduce noise and multiple point cloud data augmentation methods to increase robustness, then downsampled to decrease the computational demand followed by randomly split as training and testing sets. Experiments were then conducted to fine-tune the hyperparameters of the PointNet++ algorithm. The results showed that the optimal hyperparameter configuration entailed the adoption of the multi-scale sampling and grouping (MSG) strategy, with the augmented point cloud down sampled to 2048 points and a batch size of 8, yielding a classification accuracy of 87.7 %. The findings of our study showed the feasibility of using 3D imaging and employing the deep learning model PointNet++ for non-destructive, rapid detection of SPLCV in sweetpotato.

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.