A continuous wavelet transform-enhanced transformer model for spectroscopic detection of bacterial wilt in tomato at early infection stages.

INTRODUCTION

Tomato (Solanum lycopersicum L.) bacterial wilt is a highly destructive soil-borne disease whose early-stage infections are asymptomatic or visually indistinguishable, limiting the effectiveness of conventional diagnostic methods.

METHODS

This study developed a Continuous Wavelet Transform-enhanced Transformer framework (CWT-STNet) for early-stage detection of bacterial wilt using hyperspectral data collected from tomato canopies under controlled inoculation conditions. Infection severity was categorized into grades 0 to 4, with levels 1-2 defined as the early infection stage. Raw spectral data were fused with multi-scale wavelet features extracted using Continuous Wavelet Transform (CWT), and the effects of different scale combinations on model performance were systematically evaluated. CWT-STNet was compared with one-dimensional convolutional neural networks (1D-CNNs), random forests (RF), and support vector machines (SVM). SHapley Additive exPlanations (SHAP) were further used to interpret model predictions and identify disease-sensitive spectral features.

RESULTS

The proposed CWT-STNet framework achieved an overall accuracy of 96.77% across all infection levels and an early-stage detection accuracy of 93.60%. In comparison, models relying only on raw spectra achieved 84.15% accuracy in distinguishing early disease severity. Incorporating CWT-derived multi-scale features improved the performance of all evaluated models, while ablation analysis confirmed the importance of combining raw spectra with multi-scale wavelet features. SHAP analysis identified spectral features around 869.0 nm and the 1160-1190 nm region as particularly important for early bacterial wilt detection.

DISCUSSION: The identified spectral regions were associated with physiological and structural changes related to bacterial wilt infection, including alterations in plant water status, leaf cellular structure, and pigment-related responses. These findings demonstrate that the proposed CWT-STNet framework provides a robust, interpretable, and non-destructive approach for early detection of tomato bacterial wilt and has potential for precision disease monitoring.