Reconstructing a unified global CropAtlas for 1961-2024 via multi-modal fusion

CropAtlas is a spatially and temporally harmonized dataset providing annual physical and harvested areas for 18 major crops globally from 1961 to 2024 at 0.1° spatial resolution. The dataset covers barley, cassava, cotton, groundnut, maize, millet, oil palm, potato, pulses, rapeseed, rice, rye, sorghum, soybean, sugar beet, sugarcane, sunflower, and wheat. CropAtlas was developed through a multi-stage framework combining cropland reconstruction, crop-suitability modelling, statistical spatial disaggregation, spatiotemporal data fusion, and statistical consistency correction. First, a long-term cropland-extent dataset was reconstructed by integrating historical land-use information with 10 global cropland and land-cover products. When multiple cropland products were available for the same period, their ensemble mean was used to represent the corresponding observational field, and a three-dimensional discrete cosine transform (3D-DCT) framework was used to improve spatiotemporal consistency. Second, crop-specific ecological suitability was estimated using maximum-entropy (MaxEnt) models based on available crop-occurrence samples and environmental and socioeconomic covariates. These suitability estimates, together with reconstructed cropland extent and other spatial covariates, were used to disaggregate national and subnational annual crop statistics to individual 0.1° grid cells using the Dissever spatial-disaggregation method. Third, the statistically disaggregated estimates and 28 available global and regional crop-mapping datasets were integrated using a spatiotemporal 3D-DCT fusion framework. The method jointly processes spatial and temporal information, preserves dominant spatial structures and temporal variability at multiple scales, reduces product-specific noise and allocation artifacts, and supports reconstruction for years with limited crop-mapping observations. Crop-mapping datasets were assimilated only for the crops, regions, and years in which valid observations were available. Finally, a mass-balance correction was applied to maintain consistency between the spatially aggregated crop areas and the corresponding administrative statistics. The resulting physical-area maps represent the land physically occupied by each crop within a grid cell. Harvested areas were subsequently estimated using crop-specific planting-intensity information derived from SPAM2010 while retaining consistency with the reported annual harvested-area statistics. The two NetCDF files contain the annual global physical and harvested crop areas, respectively. CropAtlas is intended to support analyses of long-term agricultural change and may serve as a spatial input for assessments related to food security, land-use change, agricultural emissions, nutrient use, water demand, and biodiversity, as well as the development and evaluation of agricultural and land-management policies.