空氣污染仍是重大的環境威脅，而懸浮微粒是其中一項重要指標。許多研究已證實，空氣動力學直徑小於或等於 2.5 µm 的細懸浮微粒（PM₂.₅）會對公共健康造成不利影響，凸顯建立更精確估測模型的必要性。傳統數值方法在掌握空氣污染物複雜的時空變化型態方面可能存在限制，而神經網路則能從多種環境輸入變數中學習其相關性。本研究提出一種結合時間卷積網路與雙向長短期記憶網路（Temporal Convolutional Network–Bidirectional Long Short-Term Memory，TCN–Bi-LSTM）的混合模型，用於估測 2023 年逐月 PM₂.₅ 濃度。模型使用臺灣環境部自 2015 年 7 月至 2022 年 12 月的歷史空氣品質資料，並結合氣象與土地利用變數，透過 K 最近鄰－反距離加權法（k-nearest neighbors–inverse distance weighting，KNN–IDW）產製空間解析度為 1 km 的 PM₂.₅ 濃度圖。TCN 元件用於擷取時間序列中的因果型態，而 Bi-LSTM 元件則用於建模長期時間相依性。本研究的主要發現可歸納為三點。首先，所提出的模型優於所選用的傳統基準模型，包括卷積神經網路（Convolutional Neural Network，CNN）、長短期記憶網路（Long Short-Term Memory，LSTM）及隨機森林（Random Forest），其決定係數 R² 達 0.796，均方根誤差（root mean square error，RMSE）為 2.161。其次，所產製的 2023 年 PM₂.₅ 濃度圖呈現大臺北地區一致且合理的時空濃度分布型態，而季節性圖資則進一步概括了 PM₂.₅ 濃度在不同季節間的整體差異。最後，使用獨立資料集進行的補充驗證亦支持所產製圖資的可靠性。PM₂.₅ 濃度圖與太陽光度計觀測網（Aerosol Robotic Network，AERONET）及哥白尼大氣監測服務（Copernicus Atmosphere Monitoring Service，CAMS）全球再分析資料之異常相關係數分別為 0.826 與 0.882；與臺灣民生公共物聯網資料比較時，則獲得 R² 為 0.584、RMSE 為 1.975 的結果。本研究成果透過提供衛星觀測缺失或不完整區域的連續暴露表面圖，對空氣品質監測研究有所貢獻，並可降低對以氣膠光學厚度（Aerosol Optical Depth，AOD）推估 PM₂.₅ 的依賴，因為 AOD 反演常受到雲層覆蓋影響。此外，所產製的圖資亦可支援流行病學模型中的暴露評估。本研究對實務與政策具有重要意涵，提出一種利用深度學習估測 PM₂.₅ 並產製暴露表面圖的方法，可作為地面監測站的補充，尤其適用於監測覆蓋不足的地區。;Air pollution remains a major environmental threat， with particulate matter being one of its key indicators. Numerous studies have shown that particulate matter with an aerodynamic diameter of 2.5 µm or smaller (PM₂.₅) adversely affects public health， underscoring the need for more accurate estimation models. Traditional numerical methods may be limited in capturing the complex spatiotemporal patterns of air pollutants， whereas neural networks can learn correlations from multiple environmental inputs. In this study， a hybrid Temporal Convolutional Network–Bidirectional Long Short-Term Memory (TCN–Bi-LSTM) model is proposed for monthly PM₂.₅ estimation in 2023. The model uses historical air-quality data from Taiwan’s Ministry of Environment covering July 2015 to December 2022， together with meteorological and land-use variables， to generate 1-km PM₂.₅ maps through k-nearest neighbors–inverse distance weighting (KNN–IDW). The TCN component captures temporal causal patterns， while the Bi-LSTM component models long-range temporal dependencies. The key findings of the research study are threefold. First， the proposed model outperformed the selected classical baseline models: Convolutional Neural Network (CNN)， Long Short-Term Memory (LSTM)， and Random Forest， achieving an R2 of 0.796 and a root mean square error (RMSE) of 2.161. Second， the 2023 PM2.5 concentration maps produced showed coherent spatiotemporal concentration patterns across the Greater Taipei Area， while the seasonal maps further summarized broader seasonal differences in PM2.5 concentration. Finally， supplementary verification using independent datasets supported the reliability of the map products. The PM2.5 concentration maps showed anomaly correlations of 0.826 and 0.882 with AERONET and Copernicus Atmosphere Monitoring Service (CAMS) global reanalysis data， and achieved an R² of 0.584 and RMSE of 1.975 when compared with Civil IoT Taiwan data. This study’s findings contribute to air quality monitoring research by providing continuous exposure surface maps where satellite observations are missing or incomplete， thereby reducing reliance on AOD-derived PM2.5 estimates， which are often affected by cloud cover during AOD retrieval. Furthermore， the produced map products can support exposure assessment in epidemiological modelling. This study has implications for practice and policy， proposing a deep learning approach to estimate PM2.5 to produce exposure surface maps to complement ground monitoring stations， especially in areas with limited coverage.
