ABSTRACT
Neural networks provide a tool to solve many of the complex data problems and allow modeling of soil carbon and soil properties using existing data. The "Predict" software package was used to aid in processing the large data sets and in testing the numerous neural network structures. Several methods were used to avoid overfitting the training data including reducing network structure, halting training when test performance began to decline, decreasing the number of input variables, and others. If % nitrogen, and Cation exchange capacity/clay ratio data are available for other Mollisols, carbon predictions can be made with reliability. With data for other soil orders or other soil conditions, it is recommended that additional neural networks be trained to determine the best model for the specific soils. Preliminary results show good results from neural net classification of soil drainage class using soil survey data, digital elevation data, forest cover data, and SPOT imagery.
