This study resolved the problem of Part 2 and refined the method for estimating the spatial distribution of vacant houses using public municipal data by addressing the following issues.
1) This study realized a more reliable method for estimating the distribution of vacant houses without concern for the simplicity of the method used in Part 2
2) We applied the method developed in this study to the entire area of Kagoshima City and Asakura City covered by Part 2, clarifying the method’s practicality and versatility.
3) By comparing the results of Part 2 and this study, we demonstrated the extent to which the reliability of this paper’s estimation method was improved.
4) The correct answer rate on a building-by-building basis was also revealed, and the results of estimating vacant and non-vacant houses per building could be obtained.
Chapter 2 introduced the database for analyzing the characteristics of vacant houses: the vacant house database, developed by combining several public municipal data: the basic resident register (BRR), Hydrant consumption amount information (HCI), and the building registration information (BRI) and the results of a field survey of vacant houses, as used in Part 2. Chapter 3 introduced the method for estimating the distribution of vacant houses in Part 2 and its challenges.
Chapter 4 proposed a method for estimating the distribution of vacant houses. In this study, XGBoost, a decision tree-based machine learning model for dealing with missing values and setting optimal thresholds, was applied to estimate the vacant house probability of each building. XGBoost iterated the decision tree until there was no improvement in predictions, and then summed the results of each decision tree to estimate the vacancy probability per building.
Chapter 5 verified the reliability of the method proposed in this study, comparing it with the method developed in Part 2. Results show that the correct answer rate for the vacant or non-vacant judgments for each building reached 97.81% in Kagoshima City and 97.25% in Asakura City. Even when we aggregated by the 250-m square grid as in Part 2, the accuracy of this study’s methods exceeded that of Part 2. In addition, when we determined whether a building was vacant or not using the method of Part 2, the determination accuracy of vacant houses was particularly low. However, the method used in this study significantly improved this problem and could estimate both vacant and non-vacant buildings with high accuracy.
Finally, Chapter 6 showed the estimated number of vacant houses and the vacant house rate aggregated by a 500-m square grid for Kagoshima and Asakura. Compared to the results of Part 2, the estimated number of vacant houses in Kagoshima City increased from 7,361 to 9,856, while the number of vacant houses in Asakura City remained almost the same. In addition, vacant house rates were higher in the central city and mountainous areas and lower in the suburbs than in Part 2 for both cities. This indicates that it is possible to estimate the distribution of the number of vacant houses accurately.
