Version update (v3.0):A minor correction was applied to the script `predict_real_images.py` to improve the inference procedure for real experimental images. All other components, including the dataset, model architecture, and training pipeline, remain unchanged. This release contains the complete implementation and associated datasets for real-time swath width estimation in centrifugal fertilizer spreaders using a hybrid CNN-Transformer architecture. Included in this release: ResNet-18 based CNN feature extractor Transformer encoder for temporal modeling Training, validation, and inference scripts Pretrained model weights Requirements file for reproducibility Synthetic fertilizer distribution dataset Real experimental image sequences (180, 340, and 425 rpm) Labels file and documented dataset structure Dataset Description: The synthetic dataset consists of 162 video sequences per field-of-view configuration (wide FOV and narrow FOV). Each sequence contains 25 temporally ordered frames representing fertilizer particle trajectories generated based on projectile motion under gravitational and aerodynamic forces. The real dataset includes experimentally captured frames obtained under three disc speeds (180, 340, and 425 rpm), including processed frames used for inference. This version replaces the earlier example dataset and provides the full dataset required for complete reproducibility of the reported experiments. This archive is published for research transparency, reproducibility, and citation purposes.