Land use/cover change detection in East Africa using hyper-temporal biophysical and socioeconomic variables

Abstract:

Land use/cover change (LULCC) can significantly alter energy and water balance partitioning, which can affect surface and atmospheric temperatures and ultimately local and regional climate. The LULCC models used for large-area applications have limitations that may obscure the importance of land-energy-water coupling and feedbacks. Here, we present an empirical method to overcome some of these obstacles: biophysical and socioeconomic data are used to map LULCC through time on a continuous (annual) basis in Kenya. The model will be calibrated and validated using area frame samples and several geospatial datasets that are open access and available throughout sub-Saharan Africa. This ensures the approach can be used to develop LULCC maps in other parts of sub-Saharan Africa. The area frame samples have dimensions of 5 x 5 km and were developed from extensive field surveys and aerial photos taken in major agricultural regions of Kenya in 1983, 1985, 2013. A total of 2,178 frames were generated and include the percentage of 21 crop-specific land cover types and 24 additional land uses per frame, yielding a total of 45 land use/cover variables. Forty-three static and dynamic (annually-changing) predictor variables (“drivers of change”) were collected and averaged so that each area frame corresponded to a single predictor variable. Dynamic data consisted of 1) Bioclim temperature variables derived from the Princeton University Terrestrial Hydrology Group’s high resolution climate reanalysis dataset for East Africa; 2) Bioclim precipitation variables derived from the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS); 3) phenology derived from the non-linear harmonic regression of the Global Inventory Monitoring and Modeling System Normalized Difference Vegetation Index product (NDVI3g); and 4) population density derived from the United Nations Environmental Program African Population Distribution Database. Given the large number of land use/cover types and predictor variables, partial least squares regression will be used for data reduction, while decision trees will be used to generate maps of LULCC in Kenya from 1983-2013.