Comparison of two methods for increasing training

Abstract

The use of artifcial neural networks with geographical data is often constrained by the very small number of training data points which can be obtained. Using multi-spectral and parametric data from the Nullica state forest in NSW, Australia, we look at addition of noise and resampling as methods of increasing the number and quality of the training set to get the most out of the data. Resampling the data appears to offer potential as a method of generalising the neural network without the accuracy trade off of added noise.