Solução mobile baseada em visão computacional para a estimativa de parâmetros biofísicos aplicados ao monitoramento e manejo de pastagens

Abstract

Computational techniques employed in vegetation phenotyping have contributed to the development of production in the field sustainably and efficiently. In this context, the present work uses a mobile solution to obtain phenotypic data from pastures through images. For that, photos of experimental plots of grass Panicum sp were captured in two experiments. In the first experiment, biomass data were obtained from the plots, and in the second experiment, height data were obtained. The photographed pasture characteristics are extracted using computer vision, which are later correlated to the pastures’ biomass and height attributes. Three machine learning techniques were implemented: LASSO Regression, Support Vector Regression, and MultipleLayer Perceptron Network. For the selection of the best model, the Differential Evolutiontechnique was used. To compare the models, thirty repetitions of the Differential Evolution technique were performed. Evaluated using the cross-validation method, the technique that obtained the best result in both problems was the MultiLayer Perceptron Network, obtaining an average of Coefficient of Determination (R2) 0.497 for the technique best adapted to forecast biomass and 0.662 for the technique best adapted to forecast height of pasture. The results obtained indicate that the biophysical parameters of height and biomass, can be modeled in function of attributes extracted from images of pastures of the forage species Panicum maximum cv. BRS Zuri, obtained in the field. Still, software proved to be capable of geo-referencing catches and storing data.