This study aims to evaluate three different time series smoothing methods, combined or not with filtering techniques, and their impact on the agricultural land use classification, in a region of the Brazilian Cerrado, using phenological parameters extracted from Enhanced Vegetation Index (EVI) Landsat-8 image time series. We extracted the time series from pixels located on well know polygons delimited on agricultural lands and monitored on a field campaign during August 2015 and August 2016. For the classification we considered the following classes: Annual Agriculture, Natural Forest, Perennial Agriculture, Semi-Perennial Agriculture and Grassland. The three smoothing algorithms were implemented through the TIMESAT software package including the: Savitzky-Golay (SG), asymmetric Gaussian function (AG) and double-logistic function (DL), and then the phenological attributes were extracted. For each method the phenological attributes were subjected to data mining using the Random Forest (RF) algorithm. The results were evaluated by the confusion matrix analysis, including global accuracy, producer´s accuracy and kappa. The intra-class variability was measured by calculating the mean standard deviation for samples within the different classes. The best classification accuracy with the different smoothing methods was the SG applied to the raw time series, with a global accuracy of 86% and kappa of 0.82.