Assessment of Automated Image Segmentation for Predictive Soil Mapping.

Soil properties express high spatial variability, and have both random and predictable components.  This limits the vector approach of traditional soil survey. Digital soil mapping techniques and automated image segmentation provide a means to capture soil-landscape relationships that might otherwise be missed, thereby improving soil survey data.  Because soil properties are continuous across the landscape, the placement of lines separating different soil taxa approximates some critical element of either soil taxa or land management concern.  Our objectives were to assess the effectiveness of automated image segmentation for differentiation of soil taxa, soil properties, and landscape metrics in three dimensions for an order 3 soil survey.  We used Definiens Developer 7.0 to segment digital slope and surface reflectance layers for an area of southeastern Arizona's Basin and Range (~ 160,000 ha) which is currently being mapped by USDA-NRCS.  Landscape transects crossing automated segmentation lines were performed and soils were characterized in the field according to National Cooperative Soil Survey standards.  Magnitude of surface soil property variation across segmentation boundaries was quantified using standard statistical techniques.  Segmentation boundaries effectively differentiated soil map units as well as geomorphic surfaces, as indicated by slope, depth to diagnostic horizon, texture, and Munsell color.  Segmentations of Landsat band 4 and slope produced the most accurate separation of soil bodies based on visual interpretation, compared to other attempted segmentations.  Our findings suggest that slope classes and geomorphic landscape position were important for differentiating taxonomic soil classes, while surface reflectance and elevation indices proved useful for discerning individual soil properties.  This information is useful for understanding how segmentation boundaries were placed across the landscape and highlights vertical differences related to topographic indices and surface reflectance.