AI 기반 프롭테크를 활용한 토지가치의 종합적 평가

AI 기반 프롭테크를 활용한 토지가치 종합평가 방법을 제시했다.

Title: Integrated Land Value Assessment using AI-based PropTech This study presents an integrated methodology for comprehensively evaluating land value using PropTech (Property Technology), which represents the convergence of artificial intelligence and real estate technology. To overcome the limitations of conventional land value assessment that relied on quantitative variables and lacked objective measurement of environmental characteristics, this research integrates machine learning-based price prediction, text mining-based qualitative variable extraction, and lightweight transfer learning-based semantic segmentation of land cover.

This study selected public auction land as the target of analysis, given data accessibility and research ethics. While data collection for private land—including owner information, transaction records, and appraisal reports—is virtually impossible due to the Personal Information Protection Act, public auction land is publicly accessible through the Onbid system, allowing for large-scale systematic analysis. As public auction prices are actual transaction prices reflecting market supply and demand and follow the same price determination mechanism as general land transactions, the methodology developed in this study is applicable to the broader private real estate market.

First, a machine learning-based price prediction model was developed by analyzing 8,394 successful bid data of forest land sold through public auctions from 2016 to 2023. Principal Component Analysis (PCA) and Multiple Linear Regression were employed to identify key independent variables, and various algorithms, including Support Vector Regression (SVR), Random Forest, and Extreme Gradient Boosting, were comparatively analyzed. The results showed that SVR achieved the highest prediction performance with a Mean Absolute Percentage Error (MAPE) of 4.17% across bid rounds 2 through 6.

Land price change rate, number of views, number of creditors, and land area were identified as factors increasing auction prices, while landlocked status, restrictions noted in appraisal reports, and partial ownership auctions were factors decreasing auction prices. Second, an integrated prediction model was developed by extracting and refining unstructured textual information from appraisal reports of 26,365 properties (forest land, dry field, paddy field, and residential land) sold through public auctions from 2017 to 2024 using text mining techniques to construct qualitative variables, which were then combined with quantitative variables.

By applying the Instrumental Variable (IV) method and Two-Stage Least Squares (2SLS), the integrated model improved auction success probability by up to 32.08 percentage points compared to the quantitative-only analysis. Quantitative variables proved effective when historical data were abundant and economic conditions were stable, while qualitative variables (such as transportation accessibility, appraisal opinions, and legal restrictions) were more useful when data were limited or economic conditions were unstable.

Furthermore, ex-ante prediction analysis excluding auction-time information, such as bid rounds and minimum bid prices, revealed that during the stable auction market period (2017-2019), prediction accuracy of MAPE 9.74% was achieved using only appraisal report-based and economic environment variables. However, following changes in market structure induced by the COVID-19 pandemic and intensified real estate regulations (including the June 17 measures) after 2020, MAPE increased to 23-34%, confirming that public auction prices exhibited non-stationary behavior in response to market regime shifts.

Third, we performed lightweight transfer learning using pretrained FLAIR neural network models (U-Net with ResNet34 and MiTB5 backbones) provided by the French National Institute of Geographic and Forest Information (IGN). The target data comprised French IGN imagery (20 cm ground sampling distance) and Korean UAV/aerial imagery (12/25 cm ground sampling distance) provided by AI-HUB. For preprocessing, we performed class mapping between heterogeneous classification schemes and unified coordinate systems.

To improve IoU, we constructed a performance function in the form of a quadratic polynomial regression equation, using class-wise and channel-wise statistics (mean, standard deviation) of input images as hyperparameters and class-wise IoU as the dependent variable. Optimal solutions were derived through the stationary point condition of Response Surface Methodology (RSM), and the final land cover map was generated by fusing class-wise optimal results at the pixel level. Optimization was completed with approximately 60 inferences, achieving IoU improvements of up to 67.86 percentage points on Korean imagery excluded from training, compared to the baseline.

The optimized values derived from the source domain were successfully transferred to cross-regional areas between coastal and mountainous zones and cross-platform models between U-Net(ResNet34) and U-Net(MiTB5) through proportional mapping-based automatic transfer without additional inference. Statistical validation confirmed that the performance function followed a systematic quadratic response, demonstrating universal applicability across diverse geographic zones, spectral distributions, spatial resolutions, sensors, neural networks, and land cover classes.

Cost analysis revealed resource efficiency superior by 4,000-fold compared to full model training and 10–74-fold compared to prior transfer learning schemes (resource consumption per unit error reduction). This approach was executed on standard commodity CPUs. Applying the developed framework, we examined the quantitative relationship between land cover characteristics and auction price ratios for 310 parcels of paddy land auctioned in 2024, revealing that the proportion of water features exerted a statistically significant positive effect on the auction price ratio.