Classification of forest types using artificial neural networks and remote sensing data

Abstract This study classified forest types using neural network data from a forest inventory provided by the "Florestal e da Biodiversidade do Estado do Pará" (IDEFLOR-BIO), and Bands 3, 4 and 5 of TM from the Landsat satellite. The information from the satellite images was extracted using QGIS software 2.8.1 Wien and was used as a database for training neural networks belonging to the software tools package MATLAB(r) R2011b. The neural networks were trained to classify two forest types: Rain Forest of Lowland Emerging Canopy (Dbe) and Rain Forest of Lowland Emerging Canopy plus Open with palm trees (Dbe + Abp) in the Mamuru Arapiuns glebes of Pará State, and were evaluated in terms of the indicators confusion matrix, overall accuracy, the Kappa coefficient, and the receiver operating characteristics chart (ROC). The best result of classification was obtained by the probabilistic neural network of radial basis function (RBF) newpnn, with an overall accuracy of 88%, and a Kappa coefficient of 76%, showing it to be a very good classifier, and demonstrating the potential of this methodology to provide ecosystem services, particularly in anthropogenic areas in the Amazon that adopt agricultural systems with low carbon emissions.