Using fuzzy logic and spectral responses to estimate and digitally map soil organic carbon in a recently glaciated loess mantled landscape

Stephanie Frank, Purdue University

Abstract

The USDA Soil Survey has been mapping soils in the United States for nearly a hundred years. In the age of technology, high quality, continuous soil property information is needed for a variety of purposes, including policy decision making, precision agrosystems management, and environmental modeling. Due to recent global climate change research, quantification of soil organic carbon (SOC) has increased importance. Soil organic carbon estimates are currently represented as one value for a mapping unit with discrete boundaries rather than a more representative continuous map of the SOC. The purpose of this study is to develop a methodology to more statistically utilize the Soil-Land Inference Model (SoLIM). The study area is a 14 digit hydrologic unit code (HUC) watershed having an area of just over 5,260 hectares located in central Indiana that was glaciated until approximately 18,000 years ago and is part of the loess mantled, rolling glacial till plain. Several studies have demonstrated that SOC is closely related to landscape, and the SoLIM, a freely available fuzzy logic based model, was selected to create a continuous map of SOC to a depth of 50 cm. Digital elevation model (DEM) derived terrain attributes were used to determine landscape components for geomorphic positions, and study sites were selected to capture the different components. Samples were obtained to a depth of 50 cm with all sites sampled at 0-5 cm and then by genetic horizon. The SOC was determined by dry combustion and was reported as percent carbon. In previous work, conceptual models of soil formation applied by an expert soil scientist to the study area have provided the necessary input information into the SoLIM. The important terrain attributes for soil formation and their corresponding 50% membership values serve as the inputs into the SoLIM. In order to remove some of the subjectivity in soil maps, statistical methods, such as stepwise multiple linear regression and k-means clustering, combined with expert soil scientist knowledge were used to create a SOC property map. The expert user knowledge method and the stepwise linear regression method were compared and were used to select from a suite of 19 DEM derived terrain attributes those that are most important for predicting SOC. The k-means clustering algorithm was used to help partition each terrain attribute into meaningful clusters. All clusters were combined to produce meaningful soil classes. The terrain attribute cluster data for each class was used as the input in the SoLIM. Different quantile intervals were used as the 50% membership values in order to find the optimal 50% membership values. The sum squared error (SSE) was used to measure the effectiveness of each model. The model with the minimum SSE was selected as the best model for predicting SOC. A stepwise regression model using 10%/90% quantile intervals ultimately produced the best model for the area. The model provided a satisfactory map of SOC based on expected SOC distribution and SSE and provided information to further work relating SOC to landscapes.

Degree

M.S.

Advisors

Owens, Purdue University.

Subject Area

Soil sciences

COinS