Government policies to populate the Brazilian Legal Amazon induced an ongoing process of forest conversion since the 1960’s. This had big implications on climate change as enormous carbon stocks were released into the atmosphere. To understand the underlying forces driving deforestation and to assess anew carbon sequestration, knowledge of post-deforestation land use is crucial. The classification of land use requires multi-temporal remote sensing imagery to differentiate the seasonal characteristics of vegetation in target land use types. However, in the Amazon the availability of high resolution imagery is greatly constrained by persistent cloud cover, prohibiting accurate modelling of seasonal signal variations. To classify post-deforestation land use, we tested a novel approach aggregating Landsat time series stacks into temporally targeted image composites (whole agricultural year, rain season, dry season). Each composite was supplemented with 40 bands summarizing the signal variation of land use classes among a wide range of spectral indices. In order to determine the effect of different cloud masking procedures on the final classification, two independent sets of composites were produced: one processed with a single-date cloud-mask (Fmask), and the other using a multi-temporal cloud mask (Tmask). Target land use types cropland, pastures and secondary vegetation were classified employing a supervised machine learning approach (Random Forests) on temporal metrics of each composite. Better classification performance was achieved for image stacks that were produced with Tmask-derived cloud masks (overall accuracy = 88.8 %). While metrics derived from NBR time series were found to discriminate land use classes best, important predictors generally represented annual/seasonal mean and minimum values. Additionally, the classifier was tested among areas with different time lags since the last deforestation event. Although overall accuracy was lowest in areas with recent deforestation (84.1 %), results did not suggest a linear correlation between classification accuracy and the time lag to the last deforestation event. However, classification results may be strongly biased due to the strong prevalence of pastures among lag groups. Further, it was found that classification accuracy increased with additional observations, especially in the rain season. Accordingly, more accurate land use maps can be generated from data with a higher temporal frequency. The short revisit time and broad spectral coverage make Sentinel 2 a promising data source to embed the method into current forest monitoring systems.