In recent years, traffic survey methods using probe vehicles have become popular in Japan. Enormous probe vehicle data are stored in database day after day. These probe data can be utilized not only for observing individual travel behavior in a road network but also for measuring LOS (Level of Services) indexes such as travel time, loss time caused by traffic congestion, and so on. The authors have already developed a methodology of generating a digital road network by using location-positioning data of probe vehicles. This paper aims to improve the network generation methodology so that more accurate road network map can be generated. Using both hypothetical and actual data, numerical tests are examined to validate the improved methods and to estimate the data size for generating a road network with required accuracy.