Parameterizing neural network models to improve land classification performance - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 87

7
Parameterizing neural networkmodels
to improve land classification
performance
Xiaojun Yang and Libin Zhou
Neural networks are an attractive machine intelligence technique increasingly being used for pattern recognition in
remote sensing. However, the performance of a neural network is contingent upon various algorithmic and
non-algorithmic parameters. Despite significant processes over the past two decades, there is no consistent guidance
on the use of neural networks for image classification. In this chapter, we review and evaluate a set of algorithmic
parameters affecting the performance of neural networks in land classification from remote sensor data. We begin
with an introduction to the basic structure of neural networks emphasing upon the multi-layer perceptron networks
due to their robustness and popularity. Then, we discuss two focused studies we recently conducted with a satellite
imagery covering an urban area to assess the sensitivity of image classification by neural networks in relation to
various internal parameter settings and the performance of several training algorithms in image classification.
Based on literature review and our own experiments, we further propose a framework that can guide the use of
neural networks in remote sensing. Finally, we identify several areas for future research. Overall, this work can help
design better neural network models for improved performance, which can further promote the use of neural
networks as a routine tool in image classification.