UAV-based coffee yield prediction utilizing feature selection and deep learning

• Autor(a) principal: Barbosa, Brenon Diennevan Souza
• Data de Publicação: 2021
• Outros Autores: Ferraz, Gabriel Araújo e Silva, Costa, Lucas, Ampatzidis, Yiannis, Vijayakumar, Vinay, Santos, Luana Mendes dos
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/48977
• topic: Deep-learning; Remote sensing; UAV imagery; Yield prediction

Unmanned Aerial Vehicles (UAVs) combined with machine learning have a great potential for crop yield estimation. In this study, a UAV equipped with an RGB (Red, Green, Blue) camera and computer vision algorithms were used to estimate coffee tree height and crown diameter, and for the prediction of coffee yield. Data were collected for 144 trees between June 2017 and May 2018, in the Minas Gerais, Brazil. Six parameters (leaf area index - LAI, tree height, crown diameter, and the individual RGB band values) were used to develop UAV-based yield prediction models. First, a feature ranking was performed to identify the most significant parameter(s) and month(s) for data collection and yield prediction. Based on the feature rankings, the LAI and the crown diameter were determined as the most important parameters. Five algorithms were used to develop yield prediction models: (i) linear support vector machines (SVM), (ii) gradient boosting regression (GBR), (iii) random forest regression (RFR), (iv) partial least square regression (PLSR), and (v) neuroevolution of augmenting topologies (NEAT). The mean absolute percentage error (MAPE) was used to evaluate the yield prediction models. The best result was obtained by the NEAT algorithm (MAPE of 31.75%) for a reduced dataset containing only the most important features (LAI and the crown diameter) and the most important months (December 2017 and April 2018). The results suggest that a dataset of the most important month (December) could be used for the yield prediction model, reducing the need for extensive data collection (e.g., monthly data collection).