In response to the need of spatial information for supporting agricultural monitoring, we present a new remote sensing object-based approach for objective and repeatable large-scale agricultural systems mapping. This approach can be broken down into two steps: 1. A stratification of land units, based on a normalized difference vegetation index (NDVI) time series at regional level; 2. A semi-automatic land cover and land use classification, performed in each land unit at field level. To produce the land units, a principal component transformation was first applied to an annual dataset of MODIS (MODerate Imaging Spectroradiometer) NDVI images. A series of segmentations were then performed on the principal component images that contain the essential information on the physiognomy and phenology of the vegetation cover. An unsupervised evaluation method was used to identify the optimum segmentation which successfully delineates homogeneous units in terms of land use. Then, for each land unit, land cover and land use classifications were carried out at field level through segmentation of a Landsat 8 high resolution mosaic image and unsupervised classification of the MODIS NDVI time series. Finally, a Dynamic Time Warping clustering method was applied to combine the different classes by their temporal behaviour to produce a final land use systems classification which has been validated with in situ data. A map of the main cropping systems of the Brazilian state of Tocantins, an agricultural expansion region, has been successfully produced for the year 2015 following this approach. This study shows the potential of unsupervised, repeatable remote sensing approaches to provide valuable large-scale baseline spatial information for supporting agricultural monitoring.