- Demo
Through a data science project for the Finance and Taxation Bureau, Tainan City Government in December 2022, the researcher modeled the spatial distribution of location adjustment parameters with Multiple Input CNN for assessing standard values of houses in Tainan City, Taiwan.
For "POIs with a smaller range of influence" and "POIs with a larger range of influence", a grid of 100m*100m is used as the basic unit. The number of various POIs in each grid is used as the value of this grid, which CNN trained with various POI data in latitude and longitude format. For the numerical data of a single point, Linear Regression is applied to promote the performance of the prediction.
The features extracted by the three models (two CNNs, one Linear Regression) are combined, and then the fully connected layer is used to learn the feature of the location rate.
- To refine and reasonably evaluate and formulate the Location Adjustment Parameters (LAP) of the housing tax base to assess standard values of houses, it is expected to process geographic data through GIS and build a deep learning model for factor quantification to improve business efficiency and governance quality.
-
Visualize existing LAP and provide adjustment suggestions.
-
Display business premises and other facilities in the system for reference to prosperity and convenience.
-
Apply information such as the Assessed Present Value (Announced Land Price) and the Age of the Building to reflect the living value of the real estate.
-
Utilizing scientific analysis tools to improve the suitability and objectivity of operation efficiency and LAP.
Export Correlation Analysis
- Based on the existing data and the current LAP heat map, the shopping malls and the financial institutions correlate more with LAP. At the same time, the age of buildings and disgusting facilities are both negatively correlated with LAP.
｜POI
- POI Categories.
｜Land Use
- 18 types of land use status.
Data Processing
｜Numerical Data in Grid
- The data is randomly sampled, and each address is converted into latitude and longitude by TGOS and Google Map API, corresponding to the grid to which the place belongs, and the calculation is performed in 141X132 grids (18,612 in total).
｜Interpolation
- Fill in the missing LAPs in the grid. Although some areas are open spaces (such as schools, parks, etc.) without LAPs, adding them for model training and value acquisition may provide a reference for future land changes and rezoning.
｜Multiple Input CNN
-
For "POI points with small influence range" and "POI points with large influence range", a grid of 100m*100m is used as the basic unit, and the sum of various POIs in each grid is used as the value of this grid for training. Finally, the features extracted by the 3 models (two CNNs, one Linear Regression) are combined, and the fully connected layer is used to learn the LAP features.
｜Model Results & Interpretation
- In the case analysis, use the "loss between forecast and LPA in 2022" layer, and indicate the model's suggestion for the construction of the LAP of the road section:
- The redder the color, the model suggests that the LAP in this area should be increased.
- The bluer the color, the model suggests that the LAP in this area should be reduced. Model adjustment suggestions are considered reasonable and accurate after being evaluated by professionals.
- Current LAP inspection is time-consuming (including on-site survey): about 1hr/road section. Using the visual system to reduce the inspection time-consuming: about 20min/road section. Overall efficiency improvement: 67% reduction in working hours in working hours