This study investigated and comprehensively compared the performance of spectra (950–1660 nm) acquired respectively from NIR and HSI in the rapid and non-destructive quantification of azodicarbonamide (ADA) content (0–100 mg/kg) in WF and simultaneously identified WF containing excessive ADA (>45 mg/kg). The raw spectra were preprocessed using 14 methods and then mined by the partial least squares (PLS) algorithm to fit ADA levels using different numbers of WF samples for training and validation in five datasets (NTraining/Validation = 189/21, 168/42, 147/63, 126/84, 105/105), yielding better abilities of NIR Savitzky–Golay 1st derivative (SG1D) spectra-based PLS models and raw HSI spectra-based PLS models in quantifying ADA with higher determination coefficients and lower root-mean-square errors in validation (R2V & RMSEV), as well as establishing 100% accuracy in PLS discriminant analysis (PLS-DA) models for identifying excessive ADA-contained WF in each dataset. Twenty-four wavelengths selected from a NIR SG1D spectra in a 168/42 dataset and 23 from a raw HSI spectra in a 147/63 dataset allowed for the better performance of quantitative models in ADA determination with higher R2V and RMSEV in validation (R2V > 0.98, RMSEV < 3.87 mg/kg) and for discriminant models in WF classification with 100% accuracy. In summary, NIR technology may be sufficient if visualization is not required.