A neural nets urban land cover classification: A case study of brno (czechia)

Abstract

Accurate and updated land cover maps provide crucial basic information in a number of important enterprises, with sustainable development and regional planning far from the least of them. Remote sensing is probably the most efficient approach to obtaining a land cover map. However, certain intrinsic limitations limit the accuracy of automatic approaches to image classification. Classifications within highly heterogeneous urban areas are especially challenging. This study makes a presentation of multilayer perceptron (MLP), an artificial neural network (ANN), as an applicable approach to image classification. Optimal MLP architecture parameters were established by means of a training set. The resulting network was used to classify a sub-scene within ASTER imagery. The results were evaluated against a test dataset. The overall accuracy of classification was 94.8%. This is comparable to classification results from a maximum likelihood classifier (MLC) used for the same image. In built-up areas, MLP did not exaggerate built-up areas at the expense of other classes to the same extent as MLC.