Artificial Neural Network (ANN) Predictive Modeling
An ANN has been shown to be superior to other prediction methods because it can essentially learn patterns in data, unlike static methods such as multiple linear regression. Contact GIS Ag Maps if you are interested in developing an ANN predictive model or if you have questions.
Scroll down to see how ANN models, based on the same data but with different model parameter settings, can have different prediction results.
An example of how an ANN adjusts predictive ability is shown in the plots below (Hollinger, 2011). Each plot represents how accurately different ANN models predict test data (based on root mean square error [RMSE] values). Of critical importance to the success of an ANN model is how well it predicts test data. In this case, models predict crop yield based on four variables, but models can be developed for different topics with different amounts of variables. There are 5,000 different models for each of the 12 plots. The best predictive model for test data from the 60,000 total can be extracted and applied. Results are included below plots.
In the plots above, the second from the top included the model with the lowest average error, while the seventh from the top included the model with the lowest RMSE. When the two models were validated against outside data (n=19,421), multiple linear regression (MLR; using the same data) had an average error and RMSE of 4.72 and 3.95 percent higher, respectively, than the ANN model with the lowest average error and an average error and RMSE of 4.47 and 4.17 percent higher, respectively, than the ANN model with the lowest RMSE. The relationships between independent and dependent variables in the data above is very linear; ANNs can be applied for nonlinear modeling while MLR is a linear modeling method. The more nonlinear relationship become between independent and dependent variables, the better ANN should predict. Contact GIS Ag Maps for more information about ANN prediction.
Reference
Hollinger, D. 2011. Crop Condition and yield prediction at the field scale with geospatial and artificial neural network applications. Dissertation. Kent State University.