COMPARACIÓN DE MÉTODOS DE CLASIFICACIÓN APLICADOS A IMÁGENES SENTINEL-2 Y LANDSAT-8, PARA LA DIFERENCIACIÓN DE PLANTACIONES FORESTALES EN ENTRE RÍOS, ARGENTINA.

Abstract

In Argentina, the forest plantations reach 1.2 million hectares. It is considered if the image automatic classification (CA) of the sensors MSI (MultiSpectral Imager) from Sentinel-2 (S2) and OLI (Operational Land Imager) from Landsat-8 (L8) can be accurate and reliable to identify two of the most common types of plantations in the department of Concordia (Entre Ríos, Argentina), those of Eucalyptus and those of Pinus, also considering the areas already harvested. It was analyzed which combination of CA and sensor is better, and which are spatial and/or spectral characteristics of S2 and L8 that explain these differences. The study area was the department of Concordia (Entre Ríos, Argentina). Three CA methods were compared: supervised parametric (minimum Euclidean distance), supervised nonparametric (kNN) and unsupervised (Hybrid IsoData). The kNN, with an Overall Accuracy of 91.4 % for S2 is the most accurate method, mainly because it is less sensitive to the internal variability of the plots and because it is able to discriminate better two very similar categories spectrally. It is concluded that the CA is a tool with a high degree of accuracy and reliability that makes it useful to be complementary to photointerpretation. The spectral and spatial resolution of MSI does not provide a relevant improvement in the CA.