인공지능을 활용한 건축물대장 용적률 결측치 보정 모델 개발 = Development of an AI-Driven Imputation Model for Floor Area Ratio Missing Data Incorporating Parcel-Level Spatial, Temporal, and Land Use Variables

국문 초록 (Abstract)

한국 건축물대장 580만건 중 용적률 결측률이 52.7%에 달하며, 결측 시 대지면적이 96.2% 함께 누락되어 단순 대체법은 부적합하다. 본 연구는 Random Forest-INLA 통합모델로 공간·시간·용도 맥락을 반영한 적응적 대체방법을 개발했다. 대지면적 복원 가능 시 RF 단독모델이 ±10% 오차 내 70.4% 정확도를 보였으나, 복원 불가 시 RF-INLA가 RF 단독모델보다 우수했다. 공간·시간·용도 효과가 변수 중요도의 32.4%를 차지했으며, 특히 창고와 공장에서 성능이 크게 개선되었다.

다국어 초록 (Multilingual Abstract)

The National Building Registries of South Korea include 5.8 million building records, with 52.7 percent missing Floor Area Ratio (FAR)data. These missing values are likely systematic, resulting from multiple administrative digitalization processes in the 1990s, and are classifiedas Missing Not at Random (MNAR). Among the missing FAR cases, 96.2 percent also have missing data in other variables such as lot area,making simple imputation methods ineffective. This study develops adaptive imputation methods using Random Forest Integrated NestedLaplace Approximation (RF-INLA) models that incorporate spatial, temporal, and land-use factors under different data availability conditions.
For model development and validation, separate training and validation sets were used, each including 100,800 buildings stratified across 252administrative districts. When lot area data are available, a standalone Random Forest model achieves 70.4 percent accuracy within a 10percent error margin. Without lot area data, the standalone model shows a mean absolute percentage error (MAPE) of 32.9 percent, while theproposed RF-INLA model improves this to 26.8 percent. Spatial and temporal usage factors contribute to 32.4 percent of total featureimportance. The integrated approach is especially effective for industrial land uses, increasing R² by 186.6 percent for warehouses and 92.9percent for factories, with a smaller improvement of 1.6 percent for residential areas.