Recent years the cataclysm of flood has occurred in many regions around the world. For this reason, so much attention is focused on prediction of this cataclysm by creating flood risk maps and hydrodynamic – numerical simulation of flood water which are based on Digital Terrain Model (DTM). The modern techniques for automatic data acquisition provide very abundant amount of points. Actually, Light Detection and Ranging (LiDAR) is the most effective data source for DTM creation with density of one to few points per square meter and good height accuracy of less than 15 cm. This high redundancy of data is essential problem for algorithms used in programs for flood modeling. Many software generating such models are restricted with respect to the maximum number of points in DTM. Hundreds of thousands of points are too large number for complex calculations which describe fluid model of the flood water. In order to obtain reliable and accurate results, it is necessary to have DTM with an appropriate accuracy. The flood disaster also occurs in large areas what usually is associated with large data sets. However, it is possible to provide suitable DTM for flood modeling by its generalization without losing its accuracy, which could still ensure sufficient precision for hydrodynamic – numerical calculations. In this paper six reduction algorithms were tested to obtain DTM with small number of points and with accuracy comparable to the original model created from LiDAR data. The main criteria for this comparison was the relation between accuracy and reduction coefficient of final result. Methods used in this research were based on different DTM structures. GRID, TIN and hierarchical structures were compared in various approaches to obtain the most reduced and the most accurate terrain model of two study areas. As the result of the experiment the best methods for data reduction were chosen. Over 90% reduction rate and less than 20 cm root mean standard error were achieved in practice for different types of terrain with respect to input DTM. It was noted that hybrid and quadtree grid based models can be even more efficient than a typical uniform GRID or TIN one.
