以深度學習理論進行戶籍人口推估

Abstract

人口資料於各學科與領域皆有使用上需求，其中地理學注重於討論人口資料於空間上分佈位置。近年來隨著政府資料公開，Open Data可取得人口資料最精細尺度為最小統計區，但台灣政府受限於法治規定，無法開放戶籍門牌尺度人口資料。更精細的人口資料可以減少人口推估誤差，一直以來都有此需求。近年來由於電腦硬體技術提升，使深度學習理論再次受到重視與使用。近期人口推估研究也開始使用深度學習理論進行人口推估。本研究使用分區密度法，多層感知器與卷積神經網路，分別建立三種人口推估模型。並使用容積率、建蔽率、樓地板面積樓層高度、建物型態、國土利用調查成果圖、都市計畫土地使用分區圖等資料做為模型訓練因子，最終產製出5公尺人口網格資料，並與戶籍人口資料進行驗證比對。研究結果顯示卷積神經網路人口推估模型推估結果最為優秀，模型訓練表現優於多層感知器人口推估模型，卷積神經網路人口推估模型Adjusted R2可達0.72585。採用深度學習方法人口推估模型與採用傳統方法人口推估模型相比，更不容易出現極端人口高估與低估現象。

Population data are used in various disciplines and fields. Geography focuses on discussing the spatial distribution of population data. In recent years, with the disclosure of government data, Open Data can obtain population data at the most precise scale, which is the smallest statistical area. However, the Taiwan government is restricted by legal regulations and cannot open up population data at the household registration number level. There is an ongoing need for more granular demographic data to reduce errors in population estimates.In recent years, due to the improvement of computer hardware technology, deep learning theory has once again been valued and used. Recent population estimation research has also begun to use deep learning theory for population estimation.This study uses the partition density method, multi-layer perceptron and convolutional neural network to establish three population estimation models. And use data such as floor area ratio, built-up coverage ratio, floor area and floor height, building type, land use survey results map, urban planning land use zoning map and other data as model training factors, and finally produce a 5-square-meter population grid data , and verify and compare with household registration population information.Research results show that the convolutional neural network population estimation model has the best estimation results. The model training performance is better than the multi-layer perceptron population estimation model. The adjusted R2 of the convolutional neural network population estimation model can reach 0.72585. Compared with population estimation models using traditional methods, population estimation models using deep learning methods are less prone to extreme population overestimation and underestimation.