Mapping rangeland health indicators in East Africa from 2000 to 2022

Authors: SOTO, GERARDO - Austral University Of Chile; WILCOX, STEVEN - Utah State University; Clark, Pat; FAVA, FRANCESCO - University Of Milan; JENSEN, NATHAN - International Livestock Research Institute (ILRI) - Kenya; KAHIU, NJOKI - New Mexico State University; LIAO, CHUAN - Cornell University; PORTER, BENJAMIN - University Of Vermont; SUN, YING - Cornell University; BARRETT, CHRISTOPHER - Cornell University

Submitted to: Earth System Science Data
Publication Type: Pre-print Publication
Publication Acceptance Date: 11/17/2023
Publication Date: 12/1/2023
Citation: Soto, G., Wilcox, S., Clark, P., Fava, F., Jensen, N., Kahiu, N., Liao, C., Porter, B., Sun, Y., Barrett, C. 2023. Mapping rangeland health indicators in East Africa from 2000 to 2022. Earth System Science Data. https://doi.org/10.5194/essd-2023-217.
DOI: https://doi.org/10.5194/essd-2023-217

Interpretive Summary: Despite many previous mapping efforts, East Africa still lacks accurate and reliable high-resolution land cover and fractional cover classification maps necessary for management, policy, and research purposes. Machine learning classifiers and linear unmixing algorithms were applied to medium resolution (Landsat) and high resolution imagery (World View and GeoEye) available from USGS and the National Geospatial Intelligence Agency, respectively. A time-series of maps (2000-2020) at 30-m spatial resolution for land cover classes (LCC) and vegetation fractional cover (VFC, including photosynthetic vegetation PV, non-photosynthetic vegetation NPV, and bare ground BG) were created. Our products represent the first multi-decadal high-resolution datasets specifically designed for mapping and monitoring rangelands health in East Africa including Kenya, Ethiopia and Somalia thus covering a total area of 745,840 km2. These data will highly valuable for a wide range of international development, humanitarian, and ecological conservation efforts.

Technical Abstract: Tracking environmental change is important to ensure efficient and sustainable natural resources management. East Africa is dominated by arid and semi-arid rangeland systems, where extensive grazing of livestock represents the primary livelihood for most of the human population. Despite different mapping efforts, East Africa lacks accurate and reliable high-resolution land cover and fractional cover classification maps necessary for management, policy, and research purposes. Earth Observations offer the opportunity to assess spatiotemporal dynamics in rangeland health conditions at much higher spatial and temporal coverage than conventional approaches that rely on in-situ methods, while complimenting their certainty. Using machine learning-based classification and linear unmixing, this paper produced Landsat-based time series at 30m spatial resolution for mapping of land cover classes (LCC) and vegetation fractional cover (VFC, including photosynthetic vegetation PV, non-photosynthetic vegetation NPV, and bare ground BG), two major data assets to derive metrics for rangeland health in East Africa. Due to scarcity of in-situ measurements in a large, remote and highly heterogeneous landscape, an algorithm was developed to combine very high-resolution WorldView-2 and -3 satellite imagery at <2m resolutions with a limited set of ground observations to generate reference labels across the study region. The LCC analysis yielded an overall accuracy of 0.863 using our validation dataset, with Kappa of 0.841; VFC, yielded R2 = 0.801, p < 2.2e-16, normalized root mean squared error (nRMSE) = 0.123. Our products represent the first multi-decadal high-resolution dataset specifically designed for mapping and monitoring rangelands health in East Africa including Kenya, Ethiopia and Somalia, covering a total area of 745,840 km2, dominated by arid and semi-arid extensive rangeland systems. These data can be valuable to a wide range of development, humanitarian, and ecological conservation efforts and are available at https://doi.org/10.5281/zenodo.7106166 and Google Earth Engine (GEE; details in data availability section).