Applications Of Machine Learning Techniques In Predicting Selected Soil Properties In Lower Mississippi Alluvial Valley Green Tree Reservoir

Measurement of fine-scale soil properties is important for understanding and monitoring the spatial distribution of soil properties, especially in wetland areas where soil classification data are often limited. This study applied three machine learning algorithms, i.e., random forest, decision tree, and support vector machine, to predict the spatial distribution of selected soil properties (Cation Exchange Capacity, Carbon, and Nitrogen) across 12 forested wetland stands also known as green tree reservoir (GTR), encompassing 1,200 acres in the Lower Mississippi Alluvial Valley of Arkansas. A total of 540 soil samples were collected across the GTRs at a depth of 0-15cm, 15-30cm, 30-50cm. The result from this study shows that the random forest algorithms performed best in predicting CEC, carbon and nitrogen at 0-15 cm and 15-30 cm depths with decision tree providing a higher prediction accuracy for carbon at the depth of 30-50 cm. Topographic derivatives such as elevation, generalized surface TOBIA index diurnal anisotropic heating and land surface temperature were the important predictors.