應用自組織映射神經網路於環境品質模型之建構：以都市熱島因子資料為例

Abstract

本研究探討自組織映射神經網路（Self-Organizing Map, SOM）於環境品質評估的應用潛力，以都市熱島環境因子作為變數進行分群。過去環境品質評估以專家法為主，但當環境資料涉及多個面向的議題時較難以專家法訂定指數，故本研究以非監督式方法對環境資料進行分群，藉由資料本身的特徵區分不同環境樣態。相較於傳統統計方法在處理非線性與高維資料時的侷限性，以類神經網路分群方法SOM能有效映射高維資料至二維空間並保留拓樸關係，在視覺化與探索隱藏資料結構上具有更大的潛力。相似環境樣態的區塊本研究稱之為環境均質區，相較於綜合的指數，環境均質區可更直觀的應用於民眾理解及政策決策。研究區選定縣市合併前的原台中市，整合以下變數：NDVI、NDBI、NDWI、LST、建蔽率、容積率、地表粗糙度、道路密度、與工業區及大型開放空間距離等10項變數，建置SOM模型。為評估模型有效性，以Z-score大於0.5作為都市熱區範圍與分群結果進行比對驗證，並透過Random Forest（RF）進行後驗的解釋與分群合併依據。研究結果顯示SOM能有效捕捉都市相對高溫區域，以多年分LST作為變數捕捉長期熱區範圍的正確性達73%，以不同年份環境變數之環境熱區範圍正確性則介於59-66%間。 本研究以2014年SOM模型進行後處理，透過RF模型進行分群合併後以新分群作為訓練標籤，模型之Accuracy達90.5%，Macro 平均 F1-score 為 90.4%，不同分群特徵重要性不同，顯示新分群在解釋邏輯上具有差異。綜合上述，以SOM作為分群方法並結合RF於探索都市環境熱區中展現良好適用性，可作為後續綜合環境品質評估模型建構之新方向與方法論基礎。This study discusses the potential of Self-Organizing Maps（SOM）for assessing the environmental quality, and using urban heat island（UHI）related factors as a case study. Traditional assessment methods are difficult to process complex, multidimensional data. SOM is an unsupervised method that maps high-dimensional data onto a two-dimensional space. It offers advantages in visualizing and identifying hidden environmental patterns. In this study, similar environmental patterns are called environmentally homogeneous zones. Compared to traditional composite indices, representing environmental quality through homogeneous zones provides a more intuitive way for the public to understand and for policymakers to make decisions.The study area is located in Taichung, Taiwan. The study integrated ten variables: NDVI, NDBI, NDWI, building coverage ratio, floor area ratio, surface roughness, road density, and distances to industrial areas and large open spaces. Clusters were validated using a 0.5 standard deviation threshold of land surface temperature（LST）. SOM successfully identified heat-prone areas, with long-term LST-based accuracy of 73%. The annual models, using the ten integrated variables, achieved accuracies ranging from 59% to 66%.A post-analysis using the 2014 SOM model showed that reclassifying clusters with a Random Forest model yielded 90.5% accuracy and a macro-average F1-score of 90.4%. Variations in feature importance across clusters suggest distinct explanatory logic. The integration of SOM and RF is effective in identifying urban heat zones and offers a promising methodological foundation for future work.