Cooking quality of cassava roots can be assessed based on water absorption after 30 minutes of boiling (WAB30%). This study aimed to classify cassava roots into good (WAB30% = 12%) and poor (WAB30% < 12%) cooking quality using hyperspectral imaging (HSI) and machine learning. A total of 2175 samples from 121 genotypes were analyzed. After image correction and selection of regions of interest (ROIs), near-infrared average spectra were extracted. Initial data exploration showed no natural class separability, and spectral preprocessing was applied to minimize adverse effects. Due to the weak correlation between spectral variables and WAB30%, classification was preferred over direct prediction. A class imbalance issue was addressed using Synthetic Minority Over-sampling Technique (SMOTE) (Dina Elreedy, 2019), leading to 2692 samples for training and 674 for validation. The best- performing model, a Random Forest Classifier with Standard Normal Variate (SNV) preprocessing, achieved 87.54% accuracy. Pixel-wise classification maps allowed visualization of WAB30 classes distribution within samples. This study confirms the potential of Hyperspectral imaging for cassava root classification based on cooking quality. Increasing sample size significantly improved model performance, suggesting future studies should expand datasets and explore advanced learning techniques. Additionally, within-genotype variability across locations and harvests should be further investigated to refine classification models for more robust predictions.