Author: Agustina Pose
Supervisors: Felipe Lumbreras & Daniel Ponsa
Presentation time: 8:30 h
Virtual Room: 1.1 | Live presentation URL
The goal of this work is to create a set of computer vision and deep learning algorithms for automatically identifying reseed areas in crop fields, by using multispectral drone images as input. The research developed in this thesis is split in two different sections: automatic segmentation of objective vegetation (discriminating weeds and any kind of spurious vegetation) and automatic estimation of crop rows. Both problems were tackled by training a U-net, which yielded very good results. In order to overcome the lack of data and ground truth, a series of computer vision algorithms were also developed for performing an exhaustive process of data augmentation. By superimposing the obtained results (identification of vegetation / crop lines) we can identify the zones along crop rows with and without objective vegetation, being the latter the zones that should be reseeded.