過往研究中已發現土地利用與河川污染之間有著高度相關性，但目前並未有大規模的透過土地利用特徵預測河川污染物。本研究透過機器學習方法，基於內政部的土地利用資料和環境部的河川污染資訊，彌補此一空缺。資料涵蓋臺灣 54 個流域及 303 個測站，結合土地利用特徵與污染物數據進行模型訓練。研究透過因子分析篩選污染物，並應用聚類分析、隨機森林、極限梯度提升、支援向量機及人工神經網絡等模型，通過資料集切分評估準確性。結果顯示，對於特徵和污染物分群後再經由模型訓練，會比直接經由模型訓練結果更好。並且透過土地利用特徵和污染物採集年月預測污染物數值，可達到驗證集 R2（R squared）0.4 以上的結果，尤其是亞硝酸鹽氮的預測效果最佳。Shapley Value 評估指出土地利用特徵的重要性，多數人為土地利用與污染物濃度呈正相關，森林和水道沙洲灘地特徵呈負相關。本研究證明機器學習在環境污染物預測中的有效性，強調土地利用特徵的重要性，為未來污染控制和土地利用規劃提供科學依據。;Previous studies identified a high correlation between land use and river pollution, but large-scale predictions of river pollutants based on land use characteristics are lacking. This research fills this gap using machine learning, with land use data from the Ministry of the Interior and river pollution data from the Ministry of Environment. The data covers 54 watersheds and 303 monitoring stations in Taiwan, combining land use characteristics and pollutant data for model training. The study employs factor analysis to screen pollutants and uses clustering analysis, random forest, extreme gradient boosting, support vector machine, and artificial neural networks to evaluate accuracy through cross-validation. Training the model after grouping features and pollutants yields better results than direct model training. Predicting pollutant values based on land use characteristics and the collection year and month can achieve an R-squared value of over 0.4 on the validation set, with nitrite nitrogen prediction performing the best. SHAP value assessment highlights the importance of land use characteristics, indicating a positive correlation between most anthropogenic land use and pollutant concentration, while forest and river sandbank features show a negative correlation.This study demonstrates the effectiveness of machine learning in predicting environmental pollutants, emphasizing the importance of land use characteristics, and providing a scientific basis for future pollution control and land use planning.
