Purpose <p>Spatially explicit cotton fiber quality data is a key driver of crop value, yet it is rarely available to farmers compared to standard yield mapping. This lack of information limits precision management and reduces opportunities to avoid mixing low- and high-quality fiber. This study aimed to: (i) characterize the spatial variability of cotton fiber quality; (ii) develop and evaluate machine learning models for predicting fiber quality, assessing the value of temporal data and within-field heterogeneity; and (iii) determine the generalizability of these models for predicting quality across unsampled environments.</p> Methods <p>We linked georeferenced harvest data with module-level quality records from 10 commercial cotton fields in Georgia, USA, producing a total of 365 quality polygons. Random Forest (RF) models were trained using static (soil, terrain) and dynamic (multitemporal Sentinel-2 vegetation indices, VIs) predictors. The coefficient of variation (CV) of predictors captured spatial heterogeneity. Model robustness was evaluated using leave-one-field-out (LOFO) cross-validation.</p> Results <p>All eight fiber quality traits exhibited significant spatial variation. Including seasonal VIs substantially improved model accuracy, with early-season data providing the largest gains in accuracy. Including the CV further improved performance for most traits. The best-case accuracy reached R² = 0.83 for fiber reflectance. However, LOFO results showed reduced performance in unsampled fields, indicating limited generalizability without more diverse training data.</p> Conclusion <p>Cotton fiber quality varies significantly within fields and incorporating both temporal remote sensing data and within-field heterogeneity into predictive models can improve accuracy. Achieving robust field-to-field predictions will require broader training datasets, but these results demonstrate the potential for spatial quality mapping to support precision cotton management.</p>