ALGORITMOS DE APRENDIZAGEM DE MÁQUINA E VARIÁVEIS DE SENSORIAMENTO REMOTO PARA O MAPEAMENTO DA CAFEICULTURA/Machine learning algorithms and variable of remote sensing for coffee cropping mapping

• Published in: Boletim de Ciências Geodésicas 2016-10, Vol.22 (4), p.751
• Main Authors: Souza, Carolina Gusmão, Carvalho, Luis, Aguiar, Polyanne, Arantes, Tássia Borges
• Format: Article
• Language: Portuguese
• ISSN: 1413-4853; 1982-2170
• DOI: 10.1590/S1982-21702016000400043

Coffee is one of the main crops in Brazil, therefore, performing the mapping and monitoring of this culture is essential for know your special distribution. However, map this culture is not an easy task. Thus, the objective of this study was to compare the use of different variables and classification algorithms for coffee area classification. The study was conducted in three areas, environmentally different. We use 5 machine learning algorithms and 7 combinations of variables, using spectral, textural and geometric variables associated with the classification process. A total of 105 maps were made. All ratings that have not used spectral variables don't achieved good levels of accuracy. In all three areas, the algorithm that presented the best accuracies was the Support Vector Machine with overall accuracy 85.33% in Araguari, 87.00% in Carmo de Minas and 88.33% in Três Pontas. The worst results were found by Random Forest algorithm in Araguari, with 76.66% accuracy and Naive Bayes in Carmo de Minas and Três Pontas, with 76.00% and 82.00%. In all three areas, textural variables when associated with spectral, improved the classification accuracy. The SVM showed the best performance for the three areas.