Improving crop mapping in Brazil’s Cerrado from a data cubes-derived Sentinel-2 temporal analysis

Abstract

The Sentinel-2/MultiSpectral Instrument (S2/MSI) expands the frequency of satellite observations, which is relevant to elaborate detailed and timely land use and land cover (LULC) classifications. However, storing, managing, and processing big data is costly and challenging, inducing a dimensionality reduction by modeling images as composite products. Contrastingly, it obliterates the temporal resolution improvement. As LULC changes are subtle over time, little is said about how much detail we lost by degrading temporal resolution. Data cube architectures enable storing, accessing, and modeling big data, mitigating losses. Brazil Data Cube (BDC) produces multidimensional data cube collections from different medium-resolution satellite data for Brazil, including S2/MSI. Here, we evaluated three BDC S2/MSI data cubes (two 16-day composites and one unblended, with the MSI original temporal resolution) to map a dynamic-and-representative region in the far-Western Bahia agricultural belt frontier, Cerrado biome, at crop type level. We incorporate spectral indices, ground samples, and crop calendars into a Random Forest-based temporal analysis. Overall accuracies (0.91 and 0.92 for composites, and 0.96 reached for the unblended) highlight the S2/MSI temporal resolution for improving mapping tasks. Given the impact of the cropland frontier expansion over Cerrado in Brazil’s commodity production, detecting subtle landscape variations can improve agri-environmental policies.