- Soil is a complex living system with high heterogeneity, which makes locating soil map boundaries a challenge. In traditional soil survey, the placement of soil map boundaries relies largely (at least initially) on identifying the soil-forming factors of biota and topographic relief through stereo aerial photo pairs. Future soil survey is being automated with predictive digital soil mapping techniques. The blocky, raster edition of a soil map aesthetically contrasts with the smooth, visually more accurate lines of a traditional soil map. To set up rules for artificial intelligence to convert grids into smooth lines, it is essential to understand the association between the shapes of traditionally delineated soil map units and the properties used to define the map units.
In this study, taxonomically different map units with varied soil temperature regimes, soil moisture regimes, and parent materials were selected from Oregon for shape comparison. An array of shape descriptors, including Miller's circularity ratio, Schumm's elongation ratio, area deficiency ratio, full Procrustes distance, and elliptical irregularity index, were computed from soil geospatial (GIS) data to get quantitative 2-D shape descriptions for the delineations of each studied soil map unit. Kernel density estimation and two-sample Kolmogorov-Smirnov test were conducted
to evaluate the differences in shapes. Scatterplot matrix was used to display the separation of map units by linear combination of selected shape descriptors.
The purposes of this study were to i) analyze delineated 2-D shapes between taxonomically different soil map units using non-parametric statistical methods; ii) Understand if shape differences can be related to difference in soil climate regime or parent material; and iii) Find out if the linear combination of selected 2-D shape descriptors can be used to separate any pair of soil map units.
The results of this study indicate that i) delineated soil shapes appear random, but the frequency of shape archetypes is unique for each soil map unit; ii) Soil temperature and moisture regimes appear to influence such frequency, whereas parent material does not; and iii) Scatterplot matrices display separations between map units but not for all cases.
Through this study, a database containing the frequency of shape archetypes for each soil map unit can be built. This database can provide i) certain basic guidelines for machine learning, so that computer software can learn how to smooth raster grids; ii) Quality assurance and quality control (QA/QC) for traditional soil survey, so that the mapping and drawing techniques of a professional can be analyzed and learned by a novice soil mapper.
This study assessed an old question and delivered an answer foreseen but unsupported by prior work. The results of this study are more robust than earlier work, however, because this study used a more reliable database, involving a great diversity of soil orders and state factors, and advanced computer techniques and statistical analyses.