Determination of sugarcane weight under mechanized harvesting conditions is a key factor for mill harvesting operations and sugarcane ongoing research. However, frequent equipment failures in current weighing systems lead to raw material losses, higher labor time and operational costs. Existing remote sensing and sensor-based methods often require recalibration, reducing system availability and causing long weighing delays. To address these limitations, this study proposes a vision-based approach to estimate the mass of harvested sugarcane on a conveyor belt under harvesting conditions. The proposed approach combines a semantic segmentation model to isolate the conveyor belt, a Deep Neural Network (DNN) for mass estimation, and a point tracking model for speed calculation without external sensors. Training data was collected using a commercial camera, considering variations in lighting, belt color, and speed. The DNN model estimated mass with an average error of 10%, achieving optimal performance when combined with segmentation at 98% precision and an augmented dataset. This vision-based mass estimation approach could enhance mechanized harvesting efficiency by minimizing raw material losses, reducing labor time and optimizing costs in sugarcane production. Future research will explore incorporating temporal dependencies or processing shorter video segments to enhance performance.