| 摘要: | 桃園市為臺灣農地重金屬污染最嚴重之地區之一。雖經多次整治與復育，惟農地再污染風險仍未完全消除，顯示該區具有高度代表性與研究迫切性。傳統污染治理多倚賴靜態監測與事後補救，缺乏即時預警與動態預測機制，無法全面掌握污染再發生之潛勢，亦不利於智慧城市永續治理之推動。
本研究旨在整合人工智慧與資料探勘技術，建立一套應用於智慧城市農地環境管理之重金屬再污染預測模型，以協助政府部門進行風險監控、污染預防及資源配置。透過引入多元學習模型與空間決策分析，提升污染風險預測之準確性與可操作性，期能提供污染防治及風險分級決策之量化依據。
本研究以桃園市列管農地為研究對象，蒐集2004年至2021年間重金屬監測資料，選取鎘、銅與鋅三項主要污染因子。研究流程包含：（1）資料前處理與變數標準化；（2）特徵篩選與變數重要性排序；（3）模型建構與比較分析。在模型建構方面，除採用以隨機森林（Random Forest, RF）演算法外，並建構結合邏輯斯迴歸（Logistic Regression, LR）與深層神經網路（Deep Neural Network, DNN）之集成學習架構，並透過k倍交叉驗證（k-Fold Cross Validation）進行參數優化。同時結合地理資訊系統（GIS），進行污染熱區之空間視覺化分析與風險地圖繪製。結果顯示，隨機森林(RF)模型於再污染預測中表現最佳，其重金屬濃度變化預測準確率達75.76%，增量預測準確度高達99.95%；邏輯斯迴歸(LR)模型對銅、鋅、鎘之預測準確率分別為82%、83%與91%；深層神經網路(DNN)則展現高度適配性，其中鎘之預測R²值達0.98，顯示模型具高穩定性與解釋力。
本研究以灌溉小組為單位建立污染風險分級制度，並提出「灌溉小組污染預防分級管理指標」，藉由GIS繪製污染風險地圖，識別高再污染潛勢區，作為環保單位實施分級管理與監測規劃之依據。本研究之主要貢獻在於開發具實務應用性之智慧型污染預測系統，能有效降低傳統採樣與監測成本，提升政府於污染防治、農地整治與土地治理上的前瞻性決策能力。綜上所述，本研究不僅建立人工智慧導向之污染預測模式，亦落實智慧城市永續環境治理之決策支援機制，對農地復育與污染防治政策制定具重要貢獻。
;Taoyuan City is among the most severely affected regions in Taiwan in terms of agricultural heavy metal contamination. Despite multiple remediation and rehabilitation efforts, the risk of recontamination persists, underscoring both the representativeness of the region and the urgency for further investigation. Conventional pollution management primarily relies on static monitoring and post-remedial actions, lacking real-time early warning and dynamic prediction mechanisms. Such limitations impede a comprehensive understanding of potential recontamination processes and hinder the advancement of sustainable environmental governance in smart cities.
This study aims to integrate artificial intelligence and data mining techniques to develop a predictive model for heavy metal recontamination in agricultural land, designed for environmental management within the smart city framework. The proposed model assists governmental agencies in risk monitoring, pollution prevention, and resource allocation. By incorporating multiple machine learning algorithms and spatial decision analysis, the study enhances the accuracy and operational applicability of pollution risk prediction. Furthermore, the analytical results provide quantitative evidence of the key factors influencing farmland recontamination, while the model offers quantitative support for decision-making in pollution control and risk-based land management.
Focusing on regulated farmlands in Taoyuan City, this study utilized monitoring data from 2004 to 2021, targeting cadmium (Cd), copper (Cu), and zinc (Zn) as key pollutants. The analytical framework comprised three main stages: (1) data preprocessing and normalization, (2) feature selection and variable importance ranking, and (3) model development and comparative analysis. In terms of modeling approaches, this study adopts a Random Forest (RF) algorithm and an ensemble learning framework that integrates Logistic Regression (LR) and Deep Neural Networks (DNN). Model parameters were optimized using k-fold cross-validation to enhance predictive performance. In addition, Geographic Information Systems (GIS) were incorporated to perform spatial visualization of pollution hotspots and to generate risk maps. The results demonstrate that the RF model achieved the highest predictive performance, with an accuracy of 75.76% for heavy metal concentration changes and 99.95% for incremental prediction. The LR model achieved accuracies of 82%, 83%, and 91% for Cu, Zn, and Cd, respectively, while the Deep Neural Network (DNN) demonstrated strong adaptability, achieving an R² value of 0.98 for Cd, indicating high model stability and explanatory power.
A pollution risk classification system was further established based on irrigation group units, accompanied by the proposal of “Irrigation Group Pollution Prevention and Classification Management Indicators.” The GIS-based risk maps effectively identified areas with high recontamination potential, supporting environmental agencies in implementing hierarchical management and monitoring strategies. The primary contribution of this study lies in the development of a practical, AI-driven pollution prediction system that substantially reduces traditional sampling and monitoring costs while enhancing proactive decision-making in pollution control, farmland remediation, and environmental governance. Overall, this research not only establishes an AI-driven pollution prediction model but also implements a decision-support mechanism for sustainable environmental governance in smart cities, contributing significantly to agricultural land rehabilitation and pollution control policy formulation. |
