Accurate mass flow estimation of sugarcane under mechanized harvesting conditions is critical for operational efficiency in agriculture. Current approaches rely on sensor based methods, but they involve high repair costs, environmental sensitivity, and uncertain benefits. In this paper, we introduce MassFlow, a vision-based model that combines a 3D ResNet (R3D) backbone with optical flow-guided attention to capture spatiotemporal patterns under guided conditions. We propose a two-phase training approach: the first phase optimizes the alignment between flow and feature maps, and the second phase focuses on mass prediction and gradient regularization. Trained on real world videos from harvesting operations, our experiments shows that MassFlow achieves a mean absolute error (MAE) of 8.6% in per window evaluation and a mean absolute percentage error (MAPE) of 8.96 in per video evaluation, outperforming prior 2D-CNN baseline. These results demonstrate that combining spatiotemporal features with flow-guided attention improves mass estimation and reduces dependence on sensor-based systems.