Classifying soil structure using neural networks

Various feed forward artificial neural networks (ANNs) with back propagation were tested to classify 3 types of soil structure (granular, blocky, and massive) from 390 soil samples. The samples represented soils falling within the Ustoll taxonomic suborder that are part of the National Cooperative Soil Survey (USDA, NRCS) data base. The best network found to predict soil structure was one with 3 input nodes, 2 hidden nodes, and 3 output nodes with accuracies on the order of 79%. Inputs were percent organic carbon, silt, and clay. Simple perceptrons and simple linear perceptrons, e.g. networks with no hidden nodes (3 å 3), had accuracies of only 46%. Thus, ANNs are capable of learning soil structure from soil characterization data, and show a greater ability to classify soil structure types than simpler, linear methods. Classification of soil structure from commonly measured quantitative soil parameters is important because of the role structure plays in determining other soil properties, making it a critical component for modeling the soil system. This study shows the potential of artificial neural networks to recognize and learn complex relationships between quantitative soil parameters that can be used to correctly classify soil structure, and allow soil characterization data to be more effectively used for modeling activities.