Early detection of banana fusarium wilt caused by Fusarium oxysporum f. sp. cubense using hyperspectral with a metric learning strategy

BACKGROUND: Fusarium wilt of banana, caused by Fusarium oxysporum f. sp. cubense Tropical Race 4 (Foc TR4), poses a severe threat to global banana production. Early detection of this disease remains a major challenge, as infection is often widespread before visible symptoms appear. The identification of plants in the crucial asymptomatic stage is therefore paramount for effective control. To address this, we explored the potential of short-wave infrared (SWIR) hyperspectral sensing combined with deep metric learning for early-stage disease diagnosis. RESULTS: Based on hyperspectral data acquired from inoculated banana plantlets, a two-stage genetic algorithm-Shapley additive explanation (GA-SHAP) strategy was employed for band selection, yielding a compact subset of physiologically meaningful spectral bands. A hyperspectral classification framework integrating a one-dimensional convolutional neural network (1D-CNN) with an improved metric learning loss function was then developed to enhance sensitivity to subtle infection-induced changes. The proposed method achieved an average classification accuracy of 85.73% using only six selected bands. Crucially, the framework demonstrated exceptional early detection capability, achieving a diagnostic sensitivity exceeding 90%. Furthermore, spectral variations and underlying physiological mechanisms associated with Foc infection were analyzed, providing insights for scalable remote sensing applications. CONCLUSION: This study demonstrates that the integration of band selection and metric learning enables accurate and efficient early detection of Foc-induced banana wilt. The proposed framework not only offers a powerful tool for the early diagnosis of this devastating disease but also holds great promise for monitoring other plant-pathogen systems. © 2026 Society of Chemical Industry.