| Affiliation: | 1. Département de Géographie, Université de Liège, Liege, Belgium;2. International Platform for Dryland Research and Education, Tottori University, Tottori, 680-0001 Japan;3. Département de Géographie, Université de Liège, Liege, Belgium Department of Earth and Environmental Sciences, Division of Geography and Tourism, University of Leuven, Leuven, Belgium;4. CSDMS, Institute for Arctic and Alpine Research, University of Colorado at Boulder, Boulder, CO, USA;5. Independent Scholar, Edegem, Belgium;6. Arid Land Research Center, Tottori University, 1390 Hamasaka, Tottori, 680-0001 Japan;7. Department of Earth and Environmental Sciences, Division of Geography and Tourism, University of Leuven, Leuven, Belgium |
| Abstract: | Despite its environmental and scientific significance, predicting gully erosion remains problematic. This is especially so in strongly contrasting and degraded regions such as the Horn of Africa. Machine learning algorithms such as random forests (RF) offer great potential to deal with the complex, often non-linear, nature of factors controlling gully erosion. Nonetheless, their applicability at regional to continental scales remains largely untested. Moreover, such algorithms require large amounts of observations for model training and testing. Collecting such data remains an important bottleneck. Here we help to address these gaps by developing and testing a methodology to simulate gully densities across Ethiopia, Eritrea and Djibouti (total area: 1.2 million km2). We propose a methodology to quickly assess the gully head density (GHD) for representative 1 km2 study sites by visually scoring the presence of gullies in Google Earth and then converting these scores to realistic estimates of GHD. Based on this approach, we compiled GHD observations for 1,700 sites. We used these data to train sets of RF regression models that simulate GHD at a 1 km2 resolution, based on topographic/geomorphic, land cover, soil and rainfall conditions. Our approach also accounts for uncertainties in GHD observations. Independent validations showed generally acceptable simulations of regional GHD patterns. We further show that: (i) model performance strongly depends on the amount of training data used, (ii) large prediction errors mainly occur in areas where also the predicted uncertainty is large and (iii) collecting additional training data for these areas results in more drastic model performance improvements. Analyses of the feature importance of predictor variables further showed that patterns of GHD across the Horn of Africa strongly depend on NDVI and annual rainfall, but also on normalized steepness index (ksn) and distance to rivers. Overall, our work opens promising perspectives to assess gully densities at continental scales. © 2020 John Wiley & Sons, Ltd. |
