Assessment of training sample size for artificial neural networks in supervised image classification using spectral and laser scanner data

Remote Sensing techniques has gained special interest, since it can be used for monitoring systems and phenomena in local or global scale, in a temporally and spatially continuous way. Artificial Neural Networks are able to work with large amounts of data, with different characteristics. ANN was used in this work as the purpose of classifying remote sensing data. It was used multi sources and high-resolution spatial data, such as spectral images and Laser Scanner elevation data to classify the class “tree”. So, all the ANN created were specialist in tree class classification. In addition, the data used is from a densely urbanized area where there is a large variability of elevations and spectral characteristics. The results showed that the classification using both spectral and altimetry data resulted in better classifications than the use of only spectral information. It was also tested the influence of the size of samples for training the ANN, creating a learning curve for the ANN. It was noticed that with increasing the size of training samples there is a tendency to increase the accuracy in the classification. The global hits were above 87.5% when using only spectral data, and 97.5% when using spectral and altimetry data

Keywords:
Remote Sensing; Digital Image Classification; Artificial Neural Networks; High Spatial Resolution Images; Laser Scanner.