A Comparison of Supervised Segmentation Methods Based on Convolutional Neural Networks for Weed-Mapping Identification in UAV Images

Abstract

Precision Agriculture is a very important field of application, which is mainly determined by the use of high technology in agriculture. Its main goal is to increase productivity and quality, while making use of good practices to preserve the environment and at the same time optimize the use of agricultural inputs. One of the tool used in precision agriculture is the UAV, where an unmanned aerial vehicle used to image a specific area, targeting a large sampling at reduced time and costs requirement. UAV can be embedded by a light visible or multiespectral camera, allowing to identify in image several interesting patterns. One particularly useful analysis is the identification of weed, a very common kind of grass coexisting in the dominant culture. The present work proposes an comparison between state of the art convolutional neural network for identification and segmentation of weed Cynodon sp. in UAV images. Due its similarity in the visible spectrum of light, segmentation methods based on classical linear color metrics fail to properly identity the areas affected by this kind of grass. On the other hand, the use of Convolutional Neural Networks have been employed in a series of computer vision applications with success. The main goal of this work is to implement and validate the use of such convolutional approaches as a general problem-solver for weed mapping identification. The proposed approaches achieve 0.93 accuracy levels, enabling