Annual rice paddy maps of Madagascar at 10 m resolution (2017-2025)

Country-wide 10 m annual rice paddy maps for Madagascar (2017–2025), produced without ground reference labels by combining phenological pseudo-labels with Google Satellite Embeddings. This dataset accompanies the manuscript “Precise rice paddy mapping across Madagascar from 2017 to 2025 using phenological pseudo-labels and Google Satellite Embeddings.” The workflow uses the flooding-to-greenup signal from merged Harmonized Landsat Sentinel-2 (HLS) time series to generate pseudo-labels, then classifies Google Satellite Embedding (GSE) annual features with a two-stage Random Forest model refined by targeted hard samples. Files Nine annual GeoTIFFs, one per year from 2017 to 2025: MDG_rice_2nd_classified_YYYY.tif — final second-stage annual classification map. Pixel values: 1 = rice paddy, 2 = other crops, 3 = wetlands, 4 = other land covers, 255 = NoData. Format: GeoTIFF, uint8, LZW-compressed. Resolution: ~10 m. CRS: EPSG:3857 (Web Mercator). Highlights Final second-stage classifier: 91.2% overall accuracy, 99.0% precision, 83.2% recall, 0.904 F1-score (independent validation samples). Mapped paddy rice area increased from 886,112 ha in 2017 to 1,195,766 ha in 2025 (+309,655 ha, +34.9%). Most expansion occurred around existing rice-producing areas rather than distant new frontiers. A small subset of GSE dimensions consistently captured rice-relevant phenological information. Double cropping is more common in the eastern regions, while single cropping dominates much of the western and central rice area. Related resources Interactive Earth Engine App: twinsben94.users.earthengin…