Timely monitoring of corn growth at early stages is essential for food security and agricultural management, yet most existing mapping approaches depend on mature-stage spectral features, delaying operational applications. This study introduces TCBA-ViT, a hybrid framework that integrates convolutional neural networks and Vision Transformers, enhanced with dual-path Convolutional Block Attention Module (CBAM) and temporal attention, to jointly capture local spectral details and global temporal dynamics from multi-temporal Sentinel-2 imagery. Using six years (2019-2024) of data from the U.S. Corn Belt, TCBA-ViT reliably identified corn as early as June (V7 stage, four weeks after seeding) and achieved stable accuracies above 90% by late July, nearly two months before physiological maturity. Cross-year experiments demonstrated robustness to interannual variability and crop rotation, while cross-regional tests confirmed strong spatial generalization, maintaining F1-scores above 0.85 within 250 km and above 0.80 within 450 km. Compared with existing baseline models, TCBA-ViT consistently delivered earlier and more accurate classification across years and regions. Ablation analyses further highlighted the indispensable contributions of CBAM and temporal attention to performance gains. By addressing the questions of how early corn can be classified and how far models can generalize, this study provides a validated framework for early-season dynamic crop classification and large-scale agricultural monitoring, supporting sustainable decision-making.
Tan et al. (Fri,) studied this question.