Accurate mapping of winter cereals in fragmented agricultural landscapes remains challenging, as single-date spectral imagery captures only a snapshot of crop conditions, while temporal approaches alone may miss subtle biochemical and structural differences between crop types. This study proposes a hybrid framework integrating hyperspectral and phenology-informed features for winter cereal mapping in the Saïss Plain, north-central Morocco. A Spectral Attention Module (SAM) was applied to EnMAP hyperspectral imagery to identify 29 informative narrow bands. In a second scenario, these spectral features were complemented with a Dynamic Time Warping (DTW)-based dissimilarity metric derived from Sentinel-2 Enhanced Vegetation Index (EVI) time series, measuring the similarity of each parcel's seasonal profile to a reference winter cereal trajectory. Three classifiers — Random Forest (RF), Support Vector Machine (SVM), and TabPFN — were evaluated using nested random (CV) and spatial cross-validation (SCV) to assess predictive performance and spatial generalization. The hyperspectral-only scenario already achieved high separability, with spatial ROC-AUC values reaching 0.95 for SVM and 0.93 for TabPFN. Adding the DTW-based dissimilarity feature improved robustness, particularly for RF, whose spatial ROC-AUC range shifted from 0.89 (SCV)–0.98 (CV) to 0.91 (SCV)–0.99 (CV). Overall, SVM remained the strongest model in spatial ROC-AUC, while the fused spectral–temporal representation reduced low-performing cases and improved generalization under spatial validation. These findings demonstrate that combining attention-selected hyperspectral bands with a DTW-based temporal similarity metric provides an effective framework for winter cereal mapping in heterogeneous agricultural systems.