利用深度學習進行淹水影像及面積辨識

Abstract

當今受到極端氣候影響，天然災害頻傳，各國政府開始調適策略以因應氣候變遷所造成之衝擊。而臺灣位處於西太平洋颱風路徑要衝，根據中央氣象局資料統計，在過去十年間臺灣平均每年約有2.3個颱風經過；若是都市排水系統無法及時宣洩颱風所帶來的大量雨水，造成之道路淹水不僅阻礙交通，對於居民的生命財產亦造成巨大威脅。 本研究以深度學習模型為基礎，搜集臺北市區之街口CCTV (Closed-Circuit Television)影像與網際網路獲取之淹水影像，建立淹水影像辨識模式。本模式能夠準確辨識淹水期間，道路從起淹到退水時的不規則淹水；進而透過本研究所提出之面積轉換模組，將辨識出之淹水範圍轉換成推估淹水面積。水體外觀存在大量變化，在不同地點及不同氣候條件下都可能改變水體外觀；為此，參考過去文獻，證實深度學習模式在辨識圖像方面表現優異，故本研究採用近年新穎之深度學習模型Mask R-CNN (Mask Region-based Convolutional Neural Network)實施訓練；並考慮到建模時之數據集、優化器與預權重等差異性，對於淹水辨識模式準確度有一定影響，因此本研究選用兩種不同優化器，同時比較使用預權重與否及配合不同數據集，最終採用評鑑指標評估模式及最佳模式參數設定。 本研究提出模式使用Mask R-CNN搭配Adam優化器，同時以預權重取代初始權重進行建模，並以2022/08/25臺北市大安區四維路44巷CCTV影像作為案例分析。結果顯示模式在淹水標籤上的評鑑指標Precision、Recall、F1-score和IoU分別為69.5%、91.1%、0.79和81%，表現優良且在模式推估與實際淹水面積趨勢一致，顯示模式所辨識出的淹水範圍與實際淹水範圍高度相關，證明本模式適用於道路街口的CCTV。 本研究使用Mask R-CNN所建立的淹水辨識模式，能自動辨識街口CCTV影像，有效貼近淹水範圍，並透過面積轉換模組即時推估淹水面積。未來可納入水利署災害緊急應變系統，提供管理機關即時掌握是否發生淹水及淹水範圍之參考，希冀降低淹水之生命財產損失。

In the face of extreme climate impacts and frequent natural disasters, governments worldwide are adapting their strategies to cope with the repercussions of climate change. Taiwan, situated in the prime path of typhoons in the Western Pacific, faces an average of approximately 2.3 typhoons annually over the past decade according to the Central Weather Bureau. If urban drainage systems fail to swiftly manage the substantial rainfall brought by typhoons, the resulting flooding not only impedes traffic but also poses significant threats to the lives and property of residents. This study Employing a deep learning model, this study collects street intersection CCTV footage and flood images from the internet to develop a flood image detection model. This model can accurately identify irregular flooding from the onset to the receding stage, and uses an area conversion module proposed in this research to estimate the inundation area from the recognized flood extent. Given the substantial variation in the appearance of water bodies under different locations and climatic conditions, the previous literature has confirmed the superior performance of deep learning models in image detection. Consequently, this study adopts the novel deep learning model Mask R-CNN (Mask Region-based Convolutional Neural Network) for training. Considering the differences in datasets, optimizers, and pre-weights during modeling, which significantly influence the accuracy of the flood detection model, this study selects two different optimizers, compares the use of pre-weights and various datasets, and finallly uses evaluation metrics to determine the optimal model parameters. The proposed model uses Mask R-CNN with the Adam optimizer, replacing initial weights with pre-weights during modeling. The CCTV footage from Lane 44, Siwei Road, Da’an District, Taipei City, on August 25, 2022, serves as a case study. The results show that the evaluation metrics for the flood labels—Precision, Recall, F1-score, and IoU—are 69.5%, 91.1%, 0.79, and 81%, respectively, indicating excellent performance. The model's estimated flood extent closely matches the actual flood area, which demonstrates that the recognized flood range is highly correlated with the real situation and proves the model's applicability to CCTV footage at street intersections in Taipei. The flood detection model developed using Mask R-CNN in this study can automatically identify street intersection CCTV footage and effectively approximates the flood range. Using the area conversion module, it can promptly estimate the flood area in real-time. In the future, this model could be integrated into the Water Resources Agency's disaster emergency response system, and provides management authorities with real-time references on flood occurrence and extent to reduce the loss of life and property due to flooding.