低樓層建築物（透天厝、公寓住宅）為火災傷亡主要發生之建築物類型，惟火災燒大成災機率微小，但消防人力有限的前提下，如何精準找到相對危險的建築物實施防火宣導為消防機關重要課題。本研究盤點7項外顯火災危險要因，如鐵皮加蓋、加裝鐵窗、出口堆積雜物、出口停放汽車、出口停放機車、開口廣告招牌遮蔽、一樓遮雨棚等，透過消防領域專家標註2,000張高雄市低樓層建築物街景是否存在此7項要因，並運用卷積神經網路(Convolutional Neural Network, CNN)建立多標籤分類模型，以預測高雄市街景外顯性火災危險概況。
本研究透過卡方獨立性檢定發現高雄市低樓層建築物街景在「出口停放機車-出口停放汽車」、「一樓遮雨棚-出口停放機車」、「一樓遮雨棚-出口停放汽車」項目之間存在關聯。實驗結果顯示，CNN多標籤分類模型在預測表現上優於傳統問題轉換多標籤分類演算法，在0/1 Loss 、Hamming Loss、Precision@k及nDCG@k (k = 1, 2, 3, 5)評估指標上獲得較佳結果，另外，該模型表現相較CNN單標籤分類模型，在AUC、PR-AUC評估指標亦獲得更佳表現，推斷標籤間的相關性會影響模型的預測效果。

Low-rise buildings, such as houses and apartments, are the primary types of structures where fire-related casualties occur. Although the probability of a fire escalating into a major disaster is small, given the limited firefighting manpower, accurately identifying relatively hazardous buildings for fire prevention education is a critical issue for fire departments. This study identifies seven observable fire hazard factors, such as iron sheet extensions, installed iron window bars, cluttered exits, cars parked at exits, motorcycles parked at exits, obstructing billboard signs, and rain canopies. Fire safety expert annotated 2,000 street view images of low-rise buildings in Kaohsiung City to indicate whether these seven factors were present. A multi-label classification model based on Convolutional Neural Networks (CNN) was developed to predict the external fire hazard landscape of Kaohsiung City street views.
Using Chi-square independence tests, this study found correlations between certain hazard factors in the street views of low-rise buildings, particularly between "motorcycles parked at exits and cars parked at exits," "rain canopies and motorcycles parked at exits," and "rain canopies and cars parked at exits." The experimental results show that the CNN-based multi-label classification model outperformed traditional problem transformation multi-label classification algorithms across several evaluation metrics, including 0/1 Loss, Hamming Loss, Precision@k, and nDCG@k (k = 1, 2, 3, 5). Furthermore, the model demonstrated superior performance compared to CNN single-label classification models in AUC and PR-AUC metrics, indicating that label correlations play a key role in improving multi-label learning.

論文審定書 i
誌謝 ii
摘要 iii
Abstract iv
圖　次 vi
表 次 vii
第一章 緒論 8
1.1 研究背景 8
1.2 研究動機 9
1.3 研究目的 10
第二章 文獻探討 11
2.1 建築物火災風險 11
2.2 街景應用 14
2.3 多標籤分類 15
2.4 卷積神經網路 19
第三章 研究方法 20
3.1 研究流程 20
3.2 研究方法 22
第四章 實驗結果 23
4.1 資料集介紹 23
4.2 模型建立與優化 26
4.3 評估指標 27
4.4 研究結果 30
第五章 結論與未來展望 33
5.1 結論 33
5.2 研究限制 34
5.3 未來展望 35
參考文獻 36

Boutell, M. R., Luo, J., Shen, X., & Brown, C. M. (2004). Learning multi-label scene classification. Pattern recognition, 37(9), 1757-1771. doi:https://doi.org/10.1016/j.patcog.2004.03.009
Deng, M., Yang, W., Chen, C., & Liu, C. (2022). Exploring associations between streetscape factors and crime behaviors using Google Street View images. Frontiers of Computer Science, 16(4). doi:https://doi.org/10.1007/s11704-020-0007-z
Francisco CharteCharteDavid. (2015). Working with Multilabel Datasets in R: The mldr Package. The R Journal Vol. 7/2.
FuJia, T., Zhang, X., Li, S., & Zhang, Y.X.,. (2019). Do street-level scene perceptions affect housing prices in Chinese megacities? An analysis using open access datasets and deep learning. PloS one, , 14(5).
Ganda, D., & Buch, R. (2018). A survey on multi label classification. Recent Trends in Programming Languages, 5(1),, 19-23.
Gebru, T., Krause, J., Wang, Y., Chen, D., Deng, J., Aiden, E. L., & Fei-Fei, L. (2017). Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States. Proceedings of the National Academy of Sciences, 114(50), 13108-13113.
GebruKrause, J., Wang, Y., Chen, D., Deng, J., Aiden, E. L., & Fei-Fei, L.T.,. (2017). Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States. . Proceedings of the National Academy of Sciences, 114(50), 13108-13113.
Godbole, S., & Sarawagi, S. (2004). Discriminative methods for multi-labeled classification. In Pacific-Asia conference on knowledge discovery and data mining . Berlin, Heidelberg: Springer Berlin Heidelberg., 22-30. doi:https://doi.org/10.1007/978-3-540-24775-3_5
Hanley JA, McNeil BJ. (1982). The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. doi:https://doi.org/10.1148/radiology.143.1.7063747
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770-778.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25.
Law, S., Paige, B., & Russell, C. (2019). Take a look around: using street view and satellite images to estimate house prices. ACM Transactions on Intelligent Systems and Technology, 10(5), 1-19. doi:https://doi.org/10.1145/3342240
LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. doi:https://doi.org/10.1109/5.726791
LeCun, Y., Kavukcuoglu, K., & Farabet, C. (2010). Convolutional networks and applications in vision. IEEE, Proceedings of 2010 IEEE international symposium on circuits and systems, 253-256. doi:https://doi.org/10.1109/ISCAS.2010.5537907
Madjarov, G., Kocev, D., Gjorgjevikj, D., & Džeroski, S. (2012). An extensive experimental comparison of methods for multi-label learning. Pattern recognition, 45(9), 3084-3104. doi:https://doi.org/10.1016/j.patcog.2012.03.004
MadjarovKocev, D., Gjorgjevikj, D., & Džeroski, S.G.,. (2012). An extensive experimental comparison of methods for multi-label learning. Pattern recognition, 45(9), 3084-3104.
Read, J., Pfahringer, B., Holmes, G., & Frank, E. (2011). Classifier chains for multi-label classification. Machine learning, 85, 333-359. doi:https://doi.org/10.1007/s10994-011-5256-5
Tsoumakas, I. Katakis, and I. Vlahavas. (2010). Mining multi-label data. Data mining and knowledge discovery handbook,, 667-685. doi:https://doi.org/10.1007/978-0-387-09823-4_34
VandeviverC. (2014). Applying google maps and google street view in criminological research. . Crime Science,, 3, 1-16.
內政部消防署. (2024年4月). 內政部消防署全球資訊網. 擷取自 https://www.nfa.gov.tw/pro/index.php?code=list&ids=382
行政院主計總處. (2023). 縣市重要統計指標查詢系統. 擷取自 https://www.stat.gov.tw/
林元祥. (2004). 住宅建築物火災財物損失影響因素之分析與危險指標之建立. 風險管理學報, 6(1).
林盛豐委員、蔡崇義委員、楊芳玲委員. (2019). 108內調0063「20年來72名消防員因公殉職因有通盤瞭解調查之必要」調查報告. 監察院.
高雄市政府民政局. (2023年12月). 本市各區里戶口數月統計. 擷取自 https://cabu.kcg.gov.tw/Stat/StatRpts/StatRpt1.aspx?yq=112&mq=12&dq=
高雄市政府消防局. (2023). 112年高雄市消防統計年報.
陳峯鈺. (2022年7月). 火災風險導向稽查模型. 於 消防月刊. 內政部消防署.
雷明遠. (2009). 防火性能設計導入火災風險評估之研究. 內政部建築研究所.
鄭玉森. (2023). 利用事件樹以及為何樹分析探討火災案例. doi:http://dx.doi.org/10.6840/cycu202301695