The use of a multilayer perceptron for specifying the landmarks on topographic maps

Abstract

The presented article concerns the issue of landmarks selection i.e. solid objects and situational items that may be easily identified in the field.
To specify them the artificial neural networks (a multi-layer perceptron) have been used. The article describes both, how to select the most appropriate neural network architecture and input data (attribute and spatial) which are entered to the network.
The tests have been performed for the area of 4 sheets of the Military Topographic Map at 1:50 000 scale. 4 classes of objects have been analyzed (a chimney, a wayside cross, a monument and an elevation spot). To select the appropriate network architecture the cross-validation has been performed. The learning sample has been divided into 3 parts (one learning, one testing and one validation sample). This allowed to select the top 10 networks. In addition a global sensitivity analysis was conducted, which helped to determine variables with the greatest impact on the results.
Implementation of the network was made based on a test data set, located in the area of the adjacent map sheets. The results showed that the neural network was able to correctly specify a landmark. The highest index was assigned to high, isolated objects, which was in line with the way of teaching the neural network. The usage of a continuous activation function allowed to determine the index in the continuous range 0 to 1.
The spatial data from the Vector Map Level 2 and the Military Topographic Map at 1:50 000 scale have been used for studies described in this article.