Dataset of spatial products and supporting tables for the paper: "Mapping Kenya's grassland and rangeland extent using window-consistent machine learning and multi-sensor Earth observation"
Description
This dataset contains the final spatial products and supporting documentation for a national 2024 grassland/rangeland mapping study in Kenya. The products were generated using a window-consistent Random Forest classification framework that integrated manually interpreted 500 m reference image chips with multi-sensor Earth observation predictors from Sentinel-1, Sentinel-2, Dynamic World, ESA WorldCover and terrain datasets. The repository includes a 500 m continuous grassland/rangeland probability raster, the final binary grassland/rangeland extent map derived using probability threshold p ≥ 0.40 and a 5 km² minimum mapping unit, and three threshold-sensitivity products representing inclusive, minimally filtered and conservative interpretations of mapped extent. Supporting files include area-summary tables, threshold/MMU sensitivity tables, product metadata, raster inventory, data dictionary, product manifest, README documentation and MD5 checksums. The final selected binary product maps 110,277.5 km² of grassland/rangeland, equivalent to 18.8% of the valid mapped area of Kenya. The probability map is intended as the primary scientific product because it preserves uncertainty and ecological gradients, while the binary map is intended for national and regional reporting, rangeland assessment, restoration planning, biodiversity assessment, land degradation screening and land-resource applications. The dataset is not intended for parcel-level or field-boundary delineation without local validation.
Files
Steps to reproduce
Download and unzip the repository package. Review the README, metadata, data dictionary, raster inventory and product manifest in the documentation/ folder. Verify file integrity using the MD5 checksum file in checksums/. Load the main probability raster, Kenya_grassland_rangeland_probability_windowRF_500m_2024.tif, in GIS, R, Python or another GeoTIFF-compatible platform. This raster contains continuous grassland/rangeland probability values from 0 to 1. Load the final binary raster, Kenya_grassland_rangeland_binary_p40_MMU5km2_windowRF_500m_2024.tif, where 1 represents grassland/rangeland, 0 represents non-grassland/rangeland and NoData represents excluded or invalid pixels. Summing the area of pixels with value 1 reproduces the final mapped extent of approximately 110,277.5 km², equivalent to 18.8% of the valid mapped area of Kenya. Threshold-sensitivity results can be reproduced by loading the three sensitivity rasters for p ≥ 0.35 with MMU ≥ 2 km², p ≥ 0.40 with MMU ≥ 2 km², and p ≥ 0.45 with MMU ≥ 2 km², then comparing their mapped areas with Kenya_grassland_rangeland_threshold_MMU_area_sensitivity_2024.csv in the tables/ folder. The recommended final binary product is documented in Kenya_grassland_rangeland_recommended_binary_product_2024.csv.