FLUORESCENCE SPECTRAL INDICES COMBINED WITH MACHINE LEARNING CLASSIFIERS FOR EARLY SCREENING OF COCOA SWOLLEN SHOOT DISEASE

Cocoa swollen shoot disease (CSSD) is a major threat to cocoa production in West Africa, and its early detection remains difficult because current molecular methods are often destructive, costly, and time-consuming. This study investigated whether chlorophyll fluorescence spectral indices could discriminate healthy (SA, n = 361), asymptomatic infected (AS, n = 284), and symptomatic infected (SY, n = 161) cocoa leaves. Fluorescence spectra were acquired from 650 to 800 nm under 490 nm excitation, and 30 spectral indices were calculated from 806 preprocessed spectra. Their discriminatory ability was assessed using the Kruskal–Wallis test with Dunn–Bonferroni post-hoc comparisons, effect sizes, and false discovery rate correction. Nine supervised classifiers were evaluated using repeated stratified 10-fold cross-validation with fold-wise standardization. The mean spectra showed a progressive SA → AS → SY pattern, characterized mainly by increased red fluorescence around 688–691 nm and decreased far-red emission between 720 and 760 nm. All 30 indices significantly differed among classes (p_FDR < 10⁻¹¹), with the strongest effects observed for far-red descriptors, particularly F735_norm (ε² = 0.64). For three-class classification, the multilayer perceptron achieved 92.5 ± 0.7% accuracy and 90.9 ± 0.7% balanced accuracy. In the healthy-versus-asymptomatic task, it achieved an AUC of 0.991 ± 0.002, with 96.4% sensitivity and specificity. These findings indicate that chlorophyll fluorescence indices, particularly far-red descriptors, can detect pre-symptomatic spectral changes associated with CSSD and may support non-destructive field pre-screening.