Location: Dale Bumpers Small Farms Research CenterTitle: A pattern recognition approach combined with Fuzzy Logic for predicting soil properties and function. International Commission of Agricultural and Biosystems
|LIBOHOVA, ZAMIR - Natural Resources Conservation Service (NRCS, USDA)|
Submitted to: Meeting Abstract
Publication Type: Abstract Only
Publication Acceptance Date: 1/22/2018
Publication Date: N/A
Technical Abstract: Soil in the integrator of surficial processes that control the physical, chemical and biologic responses on the earth’s surface. Most soil information exists as point data or polygons with one value for soil organic carbon, soil texture, water holding capacity and bulk density to name a few. Understanding and predicting complex environmental interactions requires spatially explicit soil property predictions for model input and extrapolation of known responses. Geostatistical tools are the most common tool used to obtain continuous soil property predictions. Geostatistical tools are useful, but most areas lack the critical amount of georeferenced sampling points to predict soil differences at the field scale. The objectives of this research was: 1) to evaluate the accuracy of a hierarchal approach with pattern recognition and fuzzy logic to predict soil properties, 2) compare the new approach with traditional geostatistical tools and 3) relate to crop yield responses. This research was conducted on a 30 ha field with 60 georeferenced sampling points spaced at 30 m intervals. Soils were analyzed for texture, soil carbon and phosphorus. The georeferenced data was utilized with kriging techniques to create continuous soil property prediction at 10 m pixel resolution. The pattern recognition methodology was based on terrain algorithms of topographic wetness index, multiresolution valley bottom flatness index, ridgetop flatness index and topographic position index run in SAGA GIS. These unique terrain algorithms were groups based on common values occurring within regions of the field. The patterns were hardened and each of the 5 patterns were given unique values to represent the ideal pattern. Rules were created so that each 100% match received the ideal soil property prediction. Using fuzzy logic, all other predations were given weighted predictions based on the similarity to the pattern to create a continuous 10 m pixel resolution map. The results showed that the ordinary kriging geostatistical model performed the best giving a root mean square error of 0.25. The pattern recognition model had a root mean square error of 0.29. Even though the geostatistical model had a better statistical prediction, the pattern recognition model only relied on 5 of the 60 sample locations. The relationship to yield was more highly correlated with the pattern recognition soil property map when compared to the geostatistical map. This research has broad implications for creating continuous soil property data predictions with limited data which corresponds to biologic responses.