文章基本信息

标题：Use of Neural Networks for Land Cover Classification from Remotely Sensed Imagery
本地全文：下载
作者：L. Zhou ; X. Yang
期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2008
卷号：XXXVII Part B7
页码：575-578
出版社：Copernicus Publications
摘要：Remote sensing has been increasingly used to derive land cover information by manual interpretation or automated classification. As promising automated classifiers, artificial neural networks are difficult to parameterize, thus challenging their applicability. Previous studies investigated the impacts of some external factors, such as input data and training samples, upon neural network classification, but paid relatively little attention on how to appropriately parameterize their internal parameters. In this paper, we report the result of a pilot project aiming to develop some guides for parameterizing the multi-layer-perceptron (MLP) feed-forward back-propagation neural networks. The internal parameters we consider include the number of hidden layers, activation function, learning rate, momentum, threshold, and number of iterations. We choose part of the Atlanta metropolitan area, Georgia in the U.S.A as the test site where the landscape mosaic is characterized by several decades of rapid urban growth. We carefully configure 59 neural networks models with different internal parameters settings. We train these models with the same sample dataset and use each of them to map land cover information from a Landsat Enhanced Thematic Mapper Plus (ETM+) image. We compute the overall classification accuracy for each derived map by using the identical reference data. We found that the four internal parameters, namely, activation function, learning rate, momentum, and number of iterations, significantly affect classification accuracy; while the other two parameters, namely, threshold and number of hidden layers, also impact the classification performance. Finally, we propose a guideline that can help parameterize the MLP neural networks, and thus our study should help promote the operational use of neural networks for land cover classification
关键词：Land Cover Classification; Neural Networks; Topology; Activation Function; Neural Training; Accuracy