Abstract.A growing body of research shows the importance of land use/cover change (LULCC) on modifying the Earth system. Land surface models are used to stimulate land-atmosphere dynamics at the macroscale, but model bias and uncertainty remain that need to be addressed before the importance of LULCC is fully realized. In this study, we propose a method of improving LULCC estimates for land surface modeling exercises. The method is driven by projectable socio-ecological geospatial predictors available seamlessly across sub-Saharan Africa and yielded continuous (annual) estimates of LULCC at 5 km × 5 km spatial resolution. The method was developed with 2252 sample area frames of 5 km × 5 km consisting of the proportion of several land cover types in Kenya over multiple years. Forty-three socio-ecological predictors were evaluated for model development. Machine learning was used for data reduction, and simple (functional) relationships defined by generalized additive models were constructed on a subset of the highest-ranked predictors (p ≤ 10) to estimate LULCC. The predictors explained 62 and 65 % of the variance in the proportion of agriculture and natural vegetation, respectively, but were less successful at estimating more descriptive land cover types. In each case, population density on an annual basis was the highest-ranked predictor. The approach was compared to a commonly used remote sensing classification procedure, given the wide use of such techniques for macroscale LULCC detection, and outperformed it for each land cover type. The approach was used to demonstrate significant trends in expanding (declining) agricultural (natural vegetation) land cover in Kenya from 1983 to 2012, with the largest increases (declines) occurring in densely populated high agricultural production zones. Future work should address the improvement (development) of existing (new) geospatial predictors and issues of model scalability and transferability.Published by Copernicus Publications on behalf of the European Geosciences Union.

Abstract. A growing body of research shows the importance of land use/cover change (LULCC) on modifying the earth system. Land surface models are used to stimulate land-atmosphere dynamics at the macro- (regional to global) scale, but bias and uncertainty remain that need to be addressed, before the importance of LULCC is fully realized. In this study, we propose a method of improving LULCC estimates for land surface modelling exercises. The method yields continuous (annual) long-term (30-year) estimates of LULCC driven by socio-ecological geospatial predictors available seamlessly across sub-Saharan Africa that can be used for both retrospective and prospective analyses. The method was developed with 2252 5 × 5 km2 sample frames of the proportion of several land cover types in Kenya over multiple years. Forty-three socio-ecological predictors were evaluated for model development. Machine learning was used for data reduction and simple (functional) relationships defined by generalized additive models were constructed on a subset of the highest ranked predictors (p ≤ 10) to estimate LULCC. The predictors explained 62 % and 65 % of the variance in the proportion of agriculture and natural vegetation, respectively, but were less successful at estimating more descriptive land cover types. In each case, population density on an annual basis was the highest ranked predictor. The approach was compared to a commonly used remote sensing classification procedure, given the wide use of such techniques for macro-scale LULCC detection, and out-performed it for each land cover type. The approach was used to demonstrate significant trends in expanding (declining) agricultural (natural vegetation) land cover in Kenya from 1983–2012, with the largest increases (declines) occurring in densely populated high agricultural production zones.