Remote sensing techniques can help estimate large sugarcane areas in order to support sustainable planning and management for the sugarcane industry, as well as in the context of environmental monitoring. In this study, four generalized space-time classifiers of high-resolution satellite images were tested in São Paulo State (SP) in order to map out sugarcane areas over the course of 33 years. These classification models were created based on the random forest (RF) algorithm, using Landsat historical imagery from 1986 to 2019 as input data. Models were calibrated with reference data under the sugarcane and "other uses" classes over 10 sites across the state of São Paulo and applied to the entire state area. The four models differed in their use of input data from the different Landsat sensors. The Dice Coefficient (DC) accuracies across sites ranged between 0.60 and 0.90. The best model employed data from all sensors and was calibrated with data from 6 consecutive years. When this model was applied on independent years, the DC values averaged at 0.82. The sugarcane areas estimated from satellite imagery demonstrated a notable correlation with areas estimated from reference Canasat maps (R2 = 0.91), data from the Brazilian Institute of Geography and Statistics (R2 = 0.92), and data from the Brazilian National Supply Company (R2 = 0.73). The general trend was a large increase in sugarcane areas starting in the early 90s up until 2015, and a trend toward stabilization, and even small decreases, after 2015. This study showed that a single classification model trained on 6 years of data allowed for the accurate monitoring of sugarcane areas over 33 years, without the tedious and time- consuming work of re-calibrating the classifier for previous or new years. Year-to-year accuracy, however, largely depended on the availability and quality of Landsat imagery. The results were applicable to the monitoring of sugarcane areas over large areas and numerous years, which could a