Crop classification is a fundamental task in agricultural remote sensing. Existing methods exhibit limited generalizability due to significant phenological variations across regions and years caused by climatic differences and farming practices. This study proposes an attention-based growth-rate encoder for parcel-level satellite time-series classification. By replacing calendar time with normalized growth progress through thermal-time mapping, our method enhances model adaptability to spatiotemporal variations. Evaluated on seven crops (wheat, soybean, corn, etc.) across four regions, the approach achieved 91.22% overall accuracy and 83.47% F1-score, demonstrating superior generalization capabilities for cross-regional and cross-year scenarios.