Luís Eduardo Gonzalez Buritica, Angie Sanchez Marquina, Luis Miguel González Gutiérrez and Juan Sebastián Ochoa Cortés #
Cenicaña, Colombia
Sugarcane weight is a key indicator in sugarcane harvesting as most decision making depends on it. Currently, the weight sensor systems used in sugarcane harvesting fields are slow due to equipment malfunction and high failure rates. This work develops a vision system using a deep-learning approach to estimate the mass of harvested sugarcane on a running conveyor. Data was captured from a camera installed in the harvester conveyor under laboratory and field conditions, recording light, colors and conveyor speed variations. Sugarcane weight was estimated by training a Deep Neural Network with harvested cane load data, measured during conveyor operation to compute the accumulated mass. The removable system captures visual data and estimates sugarcane weight in real-time during the harvesting process using commercial cameras. As a result, system availability increased by 19%, average weighing time reduced by 20%, downtime reduced by 90%, and operating costs were reduced by 34%. The model that led to better results was one that was trained with dataset augmented with different techniques. A mass estimation model, coupled with a segmentation model for custom
masking and a point tracker model for speed estimation, was used to estimate mass. It leads to reduced operational costs by reducing labor time in the harvesting process and provides traceability of sugarcane production. In the future, harvested sugarcane quality and extraneous matter could be estimated to improve precision.