Key points are not available for this paper at this time.
MapBiomas is the most comprehensive land use/land cover (LULC) dataset for Brazil, widely adopted for environmental compliance (European Union Deforestation Regulation, EUDR; Brazilian Rural Environmental Registry, CAR), agricultural insurance, carbon markets, and rural credit. However, no systematic per-crop comparison against official agricultural statistics exists at municipal scale. We assess the agreement between MapBiomas Collection 10 main annual LULC product (Landsat-based, 30 m, single-class-per-pixel) and IBGE's Municipal Agricultural Production survey (PAM) across eight major crops, 5500 + municipalities, using 2023 as the reference year (last complete IBGE PAM dataset) and temporal analysis spanning 2000–2024; auxiliary MapBiomas modules (e.g., the Second Crop Agriculture module released December 2024) are outside the scope of this comparison. We identify three tiers of classification agreement: (T1) well-classified crops with near-zero aggregate gap (sugarcane: +2.9%, r = 0.96); (T2) underestimated crops where catch-all auxiliary classes recover a substantial fraction of the misclassified area (soybean: −10.4% shifts to +5.7% when Class 41 (Other Temporary Crops) is added — the overshoot itself indicating that Class 41 carries more soybean-equivalent area than is needed to close the gap; coffee: −34.5% recovers to approximately −7% with Class 48 (Other Perennial Crops), a correction effectiveness of approximately 80%); and (T3) crops requiring independent models (cotton: −65.2%, due to insufficient training samples and spectral confusion with soybean; maize: −62.8%, due to structural safrinha (second-crop) invisibility). Five distinct failure mechanisms are identified. Stratified analysis reveals a 20-fold gap variation across biomes (soybean: −1.9% in Mata Atlântica to −41.4% in Amazônia) and a frontier effect where fastest-expanding municipalities show −31.5% gap versus + 1.2% in consolidated regions. Approximately 15 million hectares of double-cropped maize are structurally invisible under annual pixel-dominant classification. These findings have direct implications for EUDR compliance verification, crop insurance exposure, and carbon market calculations. Users relying on individual MapBiomas crop classes for area estimation may face systematic disagreement of up to 65% for poorly-classified crops (e.g., cotton), while well-classified crops (e.g., sugarcane +2.9%) show aggregate disagreement near zero.
Bruno Escalhão (Sun,) studied this question.