A cross-cultivar hyperspectral framework for huanglongbing detection in citrus via wavelength optimization and deep learning.

Abstract

Huanglongbing (HLB) is a devastating disease that poses a serious threat to the global citrus industry. Due to its rapid spread and significant destructiveness, coupled with the lack of effective treatment methods, early and accurate detection is crucial for controlling disease spread and mitigating economic losses. Different citrus varieties exhibit significant differences in peel morphology and structure, leading to distinct hyperspectral reflection characteristics. This limits the applicability of traditional hyperspectral HLB detection methods across different varieties. Even when leaves with similar disease severity are detected, differences in reflection at sensitive wavelengths still exist, further limiting the adaptability of traditional hyperspectral detection methods across different varieties. To address this challenge, we propose a robust method for multi-variety HLB detection based on hyperspectral imaging. After data acquisition and preprocessing, the Successive Projections Algorithm (SPA) was used to extract characteristic wavelengths, and the Particle Swarm Optimization (PSO) algorithm was employed to identify wavelengths that remain consistent across different varieties. The statistical significance of these optimized wavelengths was verified using t-tests. The results showed that under specific conditions, there are significant differences in spectral responses among different varieties. This confirms that the selected wavelengths have cross-variety discrimination capability. Subsequently, the feature sets processed using SPA and PSO algorithms were used to train three classification models: Support Vector Machine (SVM), Multi-layer Perceptron (MLP), and a customized Convolutional Multi-scale Residual Network (CMR-CNN). The models were tested using the leave-one-out method. The PSO-optimized feature sets significantly improved model performance. In the setup where each variety was used as the test set in turn and the model was run ten times, model performance was expressed as the mean ± standard deviation (SD) of all variety test results. SVM accuracy increased from 89.38 % ± 0.81 % to 91.65 % ± 0.7 %; MLP accuracy increased from 89.58 % ± 1.19 % to 91.80 % ± 0.35 %; CMR-CNN accuracy increased from 90.83 % ± 0.43 % to 92.85 % ± 0.7 %. Due to its structural complexity and excellent feature extraction capability, the CMR-CNN model demonstrated the most outstanding diagnostic performance and showed great potential in plant disease diagnosis using hyperspectral imaging. This method establishes a universal HLB detection framework that does not require separate modeling for different citrus varieties.

Plant phenotyping relevance

柑橘葉の病徴状態を対象に、ハイパースペクトル画像の波長最適化と深層学習による品種横断的HLB検出フレームワークを開発・評価しており、植物病害表現型の取得・推定が中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.