A New Procedure for Combining UAV-Based Imagery and Machine Learning in Precision Agriculture

Abstract

Drone images from an experimental field cropped with sugar beet with a high diffusion of weeds taken from different flying altitudes were used to develop and test a machine learning method for vegetation patch identification. Georeferenced images were combined with a hue-based preprocessing analysis, digital transformation by an image embedder, and evaluation by supervised learning. Specifically, six of the most common machine learning algorithms were applied (i.e., logistic regression, k-nearest neighbors, decision tree, random forest, neural network, and support-vector machine). The proposed method was able to precisely recognize crops and weeds throughout a wide cultivation field, training from single partial images. The information has been designed to be easily integrated into autonomous weed management systems with the aim of reducing the use of water, nutrients, and herbicides for precision agriculture.