Detection of Sulfur-Fumigated Citri grandis Exocarpium Using Active Multispectral Imaging Combined with an Improved ResNet-Based Model

Rapid grading of sulfur-fumigation intensity in Citri grandis Exocarpium requires optical representations that capture both wavelength-dependent responses and heterogeneous surface morphology. This study proposes a spectral-spatial adaptive-fusion ResNet framework (SSAF-ResNet) for three-level fumigation classification. The dataset comprised 315 individual specimens, with 105 specimens each in the non-fumigated, mildly fumigated, and heavily fumigated classes, respectively. The statistical branch encoded per-band intensity distributions and local gradients, whereas the spatial branch projected the multispectral data into three contiguous spectral groups and extracted frozen ImageNet-pretrained ResNet18 features. The two PCA-RBF-SVM branch classifiers were integrated by an inner-cross-validated probability weight and evaluated by five-fold nested group cross-validation, keeping specimens imaged together within the same fold. Six classifiers under RAW, Savitzky-Golay, standard-normal-variate, and multiplicative-scatter-correction preprocessing formed the mean-gray benchmark matrix. After averaging specimen-level probabilities within each homogeneous imaging field, SSAF-ResNet achieved a field-level accuracy of 90.24 ± 5.79%, a macro-F1 of 0.895 ± 0.065, and a macro-AUC of 0.972 ± 0.018 (mean ± SD). At the individual-specimen level, the corresponding macro-F1 was approximately 0.719, highlighting the contribution of fieldwise probability aggregation. By retaining spectral distributions and spatial structure beyond mean-gray profiles, SSAF-ResNet provides exploratory evidence for homogeneous-field discrimination among controlled sulfur-fumigation treatments within a production batch.