Inductively mapping expert-derived soil-landscape units within dambo wetland catenae using multispectral and topographic data

Abstract

Constructing a cost-effective and detailed digital soil map of Africa will require the extensive utilization of both legacy soil data and legacy soil-landscape knowledge — which in Africa is primarily available from reconnaissance-scale catena or association maps and related studies. We evaluated a hybrid approach for disaggregating reconnaissance scale soil maps: rapid and inexpensive delineation of representative soil-landscape units in the field using expert knowledge, followed by the use of inductive, empirical, and correlative modeling techniques to map these landscape units. Our 2214 km 2 study area, located in central Uganda, consisted predominantly of catenae that terminate in seasonal valley floor wetlands called dambos — a type of landscape that can be found throughout the African continent. For model training and validation, we identified four landscape classes in the field using published expert knowledge: well-drained uplands (red soils); sloping dambo wetland margins (gradient > 2%), frequently inundated dambo bottoms (hummocky microtopography), and flat dambo floors (default). Using binary decision trees (BDT) with multispectral and topographic remote sensing covariates, we created a 20 m resolution map of these four classes. Multispectral inputs included reflectance values, vegetation indices, and spectral mixture modeling fractions from Système Pour l'Observacion de la Terre (SPOT) 4 satellite images acquired in December, 2006 and February, 2007. Topographic inputs consisted of a digital elevation model (DEM) from Shuttle Radar Topography Mission (SRTM) data, slope, and 20 relative elevation layers calculated using moving windows of various sizes. Decision rules were based upon the following input variables: the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Infrared Index (NDII), the shortwave infrared (SWIR) reflectance, northing, slope, and two relative elevation layers. The overall classification accuracy of 75.5% and Kappa coefficient of 0.67 suggest that a combination of multispectral, topographic, and spatial data may be used to reliably classify landscape classes for dambo-terminated catenae. At 59% of the 2214 km 2 study area, the upland class was by far the most abundant, with margins at 21%, floors at 12% and bottoms at 8%. A statistical analysis of soil property data from a small catchment located within the study area showed significant class differences in soil texture, color, organic carbon (SOC), base saturation, pH, effective cation exchange capacity (ECEC), and clay mineralogy. Though detailed soil maps are rare in Africa, reconnaissance soil maps can be inexpensively disaggregated to provide a valuable starting point for digital soil mapping.