Implementing transfer learning for citrus Huanglongbing disease detection across different datasets using neural network

Abstract

Citrus Huanglongbing (HLB) is highly contagious, and timely detection and removal of HLB-infected citrus trees is extremely important to prevent its spread. However, the robustness of optical imaging-based models remains limited by the variations in data due to different plant varieties, geospatial conditions, and data collection dates, etc. This study aimed to propose a method for robust HLB detection via transfer learning with multispectral-multicolor imaging. Four lightweight neural networks, namely Yolov7, Yolov7-tiny, Yolov4-tiny, and Mask-RCNN were introduced for citrus HLB disease detection across different datasets. Transfer learning on the Orah mandarin dataset was conducted using the Navel orange dataset for pre-training. The results showed that Mask-RCNN achieved the best performance with an mAP@0.5 of 91.65%. By replacing the backbone of Mask-RCNN with MobileNetV3-large, the model Mask-RCNNV3 was established, with an mAP@0.5 of 93.37% and then used for transfer learning for other datasts. Further optimizing the number of transferred layers and sample size, it revealed the most favorable sample size was 20 per class, and the mAP@0.5 gradually increased at the first 9 layers. Mask-RCNNV3 under the best transfer learning parameters, called Mask-RCNNV3_best, achieved the mAP@0.5 of 93.14% for Orah mandarin, 91.82% for Blood orange and 92.36% for Ponkan, respectively. Compared to the original Mask-RCNN model, the training parameters (Params) and GFLOPs were reduced by 82.95% and 96.57%, respectivley. It demonstrated that a limited amount of labeled data proved sufficient to achieve satisfactory performance across the tested cultivars and growing conditions. The FPS of the model was also improved by 4 times compared to Mask-RCNN, illustrating the potential of the model for edge deployment for practical applications. These findings would bridge the gap between research and practical implementation, reduce costly labeling for model training and provide practical tools for citrus growers to use.

Plant phenotyping relevance

柑橘HLB感染状态をマルチスペクトル画像から推定するニューラルネットワークを開発・比較・転移学習で検証しており、植物病害フェノタイピング手法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.