Optimal path/neutral [i.e. neural] network approaches to modeling of forest road design for use in automated GIS systems

A model that integrates the use of feed-forward Artificial Neural Networks and GIS techniques was developed for preliminary road design in forestry. The model was built and tested with GIS data from Malcolm Knapp Research Forest (approximately 5000 ha). Artificial Neural Networks having a variety of architectures were trained using this data, in conjunction with a several different learning parameters. Once the neural networks were trained and tested, the knowledge was used to predict based on new data. The results of these predictions consisted in a numerical representation of the "likelihood" of a given cell to contain a road. The resulting values were represented in GIS format as a new cost - surface. With the addition of this new surface, an optimal path function was run to trace the location of the road. The Artificial Neural Network approach was compared to the classical "Slope - Curvature" approach, traditionally used before in preliminary forest road design with GIS. In all cases, the Artificial Neural Network approach yields better results, both in terms of road length and spatial autocorrelation coefficient. The results of this study suggest that the technique used may be a reasonable method for determining the preliminary road location in forested areas. The roads generated with the aid of Artificial Neural Networks have a higher spatial autocorrelation with the original road and are also shorter than the ones generated with the Slope - Curvature cost surface. The route selected is not perfect from an engineering point of view; however it still can provide good information for the forest road-planner. The route selected by the computer should be modified based on the field inspection and the knowledge and experience of the specialist.Forestry, Faculty ofGraduat