- 學位論文
多層多類分區密度之空間人口重分布模式
A Multi-Layer and Multi-Class Dasymetric Model for Reconstructing Spatial Population Distribution
摘要
詳細正確的人口資料與人口的空間分布是相關決策的重要依據，但因涉及隱私保護因素，一般多將人口資料加總統計後再以行政區界發布，易產生分區單元不夠細緻、空間型態扭曲或因行政邊界變動無法進行跨時間的時序性分析等問題。且於災害發生時，因受災範圍與行政單元邊界不一致或橫跨數個行政區，不易得知正確的受災人數，雖可用面積內差法推估受災人數，但實際人口分布與面積內差下的區域之均勻分布假設有所矛盾，故如何合理重分配加總的人口資料是極為重要的課題。 本研究所建立之多層多類分區密度(multi-layer multi-class dasymetric model, MLMCD)的人口重分配模式，是以加總的人口資料為基礎，重建人口空間分布趨勢，可以合理將大區域（如縣市）加總之人口資料重分配至小網格單元。MLMCD是根據輔助資料進行分層與分類，層間具有從屬關係，層內有類別之分類，需要以社經資料決定各類別之人口重分配權重。研究中以臺北縣、市為例，將縣市的總人口重新分配至網格，利用統計區分類系統中的村里、最小統計區與網格之空間單元進行誤差比較，除了以統計的絕對平均誤差(mean absolute percentage error, MAPE)評估人口重分配的正確性外，進一步利用誤差矩陣(error matrix)及Kappa指標比較人口重分配後的空間分布與原始資料的空間分布是否一致，用來評估MLMCD人口重分配之正確性。 研究顯示面積內差法與MLMCD的人口推估結果，在山區呈現被高估現象，平地區域為低估，但MLMCD之高估與低估幅度較小並具有逐層改善誤差特性存在。以誤差的統計指標而言，由第零層至第三層均有逐層改善的結果：MAPE值分別由0.99降至0. 13(以網格單元為比較基準)；0.866降至0. 583(以最小統計區單元為比較基準)；0.809大幅降至0.458(以村里單元為比較基準)。有關空間分布型態之差異分析上，Kappa值由0.351提升至0.814（以網格為比較基準）;由0.669提升至0.888(以最小統計區為比較基準)，顯示以MLMCD進行人口重分配的結果與原資料之空間分布型態具有一致性。
並列摘要
Detailed and correct spatial distributions of population are the foundation for sound regional planning and management decisions. Population data are usually disseminated in aggregated form for confidentiality concerns. However, this approach of spatial aggregation may distort the original spatial pattern from the modified areal unit problem (MAUP). And the frequent changes of boundaries over time may make the across temporal analysis impossible. It is difficult to estimate population at risk in disaster because boundary of the disaster area may not coincide with the population aggregation units. Although population at risk may be estimated using areal interpolation method, errors may arise from unreasonable assumption of uniform distribution in the aggregation area. Effective algorithms to disaggregate the aggregated population into smaller spatial units get more and more important. A Multi-Layer Multi-Class Daymetric (MLMCD) model was developed in this study to reconstruct spatial distributions from spatially aggregated population data. Ancillary data, such as remote sensing imageries, census, land use, traffic network and other infrastructure were used to disaggregate the aggregated population data into smaller grids. These disaggregated grid data were then summed up to different spatial levels for error comparisons. Mean Absolute Percentage Error (MAPE) was used to examine the effectiveness of the proposed MLMCD model in this population disaggregation process. Error matrix and Kappa Index as in remote sensing were used to compare the spatial distribution pattern using hotspot analysis. From the case study in Taipei metropolitan area, the results show the error is decreased as the layer increased and more ancillary data were used. The MAPE are significantly improved from layer 0 to layer 3. MAPE decreased from 0.99 to 0.13 (compared at grid level), from 0.866 to 0.583 (compared at census tract level) and from 0.809 to 0.458 (compared at Li administration level). Besides, the increases of Kappa indices from 0.351 to 0.814 (at grid level) and from 0.669 to 0.888 (at census tract level) shows that the proposed MLMCD model effectively preserve the spatial distribution characteristics of population in the disaggregation process.