Document Type

Thesis - Open Access

Award Date


Degree Name

Master of Science (MS)

Department / School

Agronomy, Horticulture, and Plant Science

First Advisor

Douglas Malo


Soil organic carbon (SOC) is an important soil parameter of cultivated soils that needs to be monitored and mapped regularly to enhance soil health and productivity. SOC levels in cultivated areas is difficult to monitor for farmers and is costly to analyze using traditional methods. The objective of this study was to estimate surface SOC distribution in selected soils of Major Land Resource Areas (MLRA) 102A (Rolling Till Plain, Brookings County, SD) and 103 (Central Iowa and Minnesota Till Prairies, Lac qui Parle County, MN) using soil sample data, Web Soil Survey (WSS) data, and satellite imagery (Landsat 8 and PlanetScope). Different satellite imagery bands and band combinations were used to reach more accurate results. The dominant soils in the area are Haplustolls, Calciustolls, and Endoaquolls formed in silty sediments, local silty alluvium, and till. Sites were selected and soil samples were collected in May 2018 after planting. SOC and soil properties were measured at the 0-15 cm depth. SOC was mainly affected by soil texture in the studied selected soils. Multiple-linear regression was used to build SOC prediction models from soil test data. The final SOC model (using stepwise regression) is SOCp = 3.98 + (-0.210 pH) + (-0.220 Sand [g kg-1]) + (0.040 Sum of Extractable Cation, SOEC [cmolc kg-1]). The Ridge Regression (RR) (CV = 0.066, MSE = 0.063) and Principal Component Regression (PCR) (CV = 0.071, MSE = 0.068) were used to deal with multicollinearity and RR was determined to be as the best model, with 82.7% of variation in SOC explained by the RR model. Landsat 8 and PlanetScope spectral bands and different indices were also used to develop SOC prediction models. The stepwise regression analyses revealed that the Landsat 8 prediction model had multicollinearity problem. Ridge regression and PCR were applied, and RR was chosen as the best model with SOCp = -26.7 + (0.310 BSIL) + (-23.2 Band 5L) + (75.8 Band 2L) + (-51.1 Band 3L) + ( -3.05 Band 7L). The RR model (CV = 0.24, MSE = 0.22) explained 37.0% of the variation in SOC for Landsat 8. The reduced PlanetScope model was SOCp = -25.1 + (2980 Band4P) + (0.327 BSIP). Approximately 60.0% of the variation in SOC was explanined by the Ordinary Least Square (OLS) (CV = 0.15, MSE = 0.14) model and was free of multicollinearity. WSS data showed similar patterns as soil test data for SOC predictions. The best model for WSS data was a linear regression, SOCp = 3.37 + (-0.0200 Sand WSS [g kg-1]) and 49.0% of the variation in SOC was explained by this model. WSS data were then added as variables into the spatial (satellite) estimation models. The Landsat 8 and WSS data explained 53.3%, PlanetScope and WSS data explained 68.8% of the SOC variation. Based on these results, deciding on the number of soil sampling points, and the use of specific variables in the model is very crucial for the model development. Estimating SOC by minimizing the number of needed soil sampling points, using satellite imagery, and public free sources provides an easy, efficient and cost-effective way to monitor SOC levels and identify the best management systems for producers and natural resource managers. This project produced accurate SOC prediction models using soil test data, satellite imagery and Web Soil Survey data. This SOC estimation model helps farmers, resource managers, and researchers to monitor SOC concentration on the soil surface using remote sensing alone, or with WSS data, or with a minimal amount of soil test data.

Library of Congress Subject Headings

Soils -- Organic compound content -- South Dakota.
Soils -- Organic compound content -- Minnesota.
Soils -- Carbon content -- South Dakota.
Soils -- Carbon content -- Minnesota.
Soils -- Remote sensing.




South Dakota State University



Rights Statement

In Copyright