Non-destructive detection of Huanglongbing using hyperspectral reflectance and machine learning algorithms

Huanglongbing (HLB), caused by the pathogen Candidatus Liberibacter asiaticus, is one of the most destructive diseases affecting citrus cultivation worldwide.
Once the pathogen is successfully transmitted to a tree by citrus psyllids, the tree will eventually die, as no cure has been established.
Thus, early detection and removal of infected trees is one of the most essential elements of the current HLB management to prevent further expansion of this disease.
Here, we report a non-destructive field HLB detection method based on hyperspectral reflectance data of Citrus reticulata. A total of 72 hyperspectral data sets of leaves were collected in an orchard severely affected by HLB in Thailand.
A reported high precision machine learning algorithm random forest was employed to distinguish between visually HLB symptomatic and asymptomatic leaves.
We used principal component analysis (PCA) to extract 16 key feature bands for HLB detection and evaluated all combinations by an incremental feature selection method.
A combination of nine bands (727, 930, 933, 936, 939, 942, 954, 957 and 997 nm) achieved an F1-score of 99.83%, showed that random forest using PCA was effective in analyzing hyperspectral data.
We then utilized nearest neighbor matching method to evaluate the reliability of our expert system and the results showed that the feature bands identified by expert visual inspection and PCR test differed by only 2.78 nm, which is well within the 7 (0±3.5) nm spectral resolution of the hyperspectral camera.
This suggests that our expert system effectively identifies HLB features, even those invisible to the human eye.

XV International Citrus Congress

R. Dong, A. Shiraiwa, K. Ichinose, A. Pawasut, S. Mensin, M. Sumalee, T. Hayashi

Citrus reticulata, disease detection, feature extraction, hyperspectral imaging, random forest

https://doi.org/10.17660/ActaHortic.2026.1448.89