World-Maize-10: a global 10m maize distribution dataset derived from Sentinel-2 and Sentinel-1 imagery from 2021 to 2024

High-resolution, temporally consistent global maize maps are essential for food-security assessment and supply-chain risk analysis. However, existing products are limited in terms of spatial resolution, temporal continuity, and transferability across regions, primarily due to the scarcity and inconsistency of reliable training samples in different climates and management regimes. We present a knowledge-guided, sample-efficient global maize mapping framework that can distinguish between main- and second-season maize. The framework converts the distinctiveness of maize during the peak season into transferable labels by integrating grid-specific peak-window screening with optical-radar consistency checks throughout the growing season. For each 2°×2° grid, we derived an adaptive time window based on maize peak and phenology priors, and removed pixels outside of off-phenology or low-greenness to stabilize the candidates. We then designed two peak-anchored indices, the Maize Spectral Index (MSI) and the Maize Radar Index (MRI), and developed a novel Peak and Temporal Weighted Dynamic Time Warping (PTW-DTW) algorithm to suppress maize-like confounders and generate spatially coherent, high-confidence maize samples that can be reused across regions and years. Training grid-adaptive Random Forest classifiers produced World-Maize-10: the first annual, global maize datasets at 10 m resolution from 2021 to 2024 that explicitly map both the main and second growing seasons. Independent, pixel-level validation demonstrated stable performance across years, achieving an overall accuracy exceeding 91% for main-season maize and approximately 90% for second-season maize, as verified by in-situ samples. At the national and subnational scales, the mapped areas showed strong agreement with official statistics (R²≈0.99) and remained consistent across major producing countries. Overall, our framework provides a scalable, update-ready, open-access global mapping pipeline, as well as the annual 10m maize products that address persistent limitations in training data and support agricultural monitoring on a global scale.The World-Maize-10 dataset is archived on the public repository Figshare. The product is provided as global 10m binary raster tiles, where pixel value 1 denotes maize and 0 denotes non-maize. File names follow the convention Maize_Season_Year_GridID.tif, where Season indicates the main or second season, Year is the mapping year, and GridID identifies the mapping grid. To facilitate spatial filtering and tile selection, we also provide two grid shapefiles uploaded together with the raster products: Main_season_grids.shp and Second_season_grids.shp. The attribute fields Grid_main and Grid_sec represent the grid IDs for the main-season and second-season mapping grids, respectively. Users may first intersect their area of interest with these shapefiles to select the relevant grids, and then download the corresponding raster tiles.We gratefully acknowledge ESA for providing Sentinel-1/2 data and the GEE platform for computational support. We also thank the WorldCereal project for its pioneering efforts in global maize mapping and for establishing an open, community-based reference data repository, as well as the many contributors worldwide who have shared field and reference data over the years; together, these efforts laid an essential foundation for the conception and validation of this study. We are indebted to our team members and collaborators for their support in collecting and curating ground reference data.