Robotic Fertilization in Strip Cropping using a CNN Vegetables Detection-Characterization Method

Abstract

To meet the increased demand for organic vegetables and improve their product quality, the Sureveg CORE Organic Cofund ERA-Net project focuses on the benefits and best practices of growing different crops in alternate rows. A prototype of a robotic platform was developed to address the specific needs of this field type at an individual plant level rather than per strip or field section. This work describes a novel method to develop robotic fertilization tasks in crop rows, based on automatic vegetable Detection and Characterization (D.a.C) through an algorithm based on artificial vision and Convolutional Neural Networks (CNN). This network was trained with a data-set acquired from the project’s experimental fields at ETSIAAB-UPM. The data acquisition, processing, anc actuation are carried out in Robot Operating System (ROS). The CNN’s precision, recall, and IoU values as well as characterization errors were evaluated in field trials. Main results show a neural network with an accuracy of 90.5% and low error percentages (<3%) during the vegetable characterization. This method’s main contribution focuses on developing an alternative system for the vegetable D.A.C for individual plant treatments using CNN and low-cost RGB sensors.