人工智慧技術於即時淹水影像辨識之研究

Abstract

受全球氣候變遷加劇影響，短延時強降雨之發生頻率及強度日漸嚴重，連帶增加淹水災害發生之可能性，對於城市和地下水道之基礎設施構成了嚴峻挑戰，而淹水發生區域及範圍定位資訊對於即時搶救災至關重要。傳統上，淹水事件多仰賴人力進行通報或於淹水熱區建置路淹感測器，然而這些作法缺乏時效性且需考量裝設成本及維護管理，亦僅能提供區域性之監測數據。因此，如何有效且具即時性之淹水檢測方法乃為刻不容緩之議題。 本研究基於深度學習模型開發即時淹水影像辨識模式，並透過路口監視器影像作為資料輸入，盼以既有廣泛佈設之CCTV (closed-circuit television)取代路淹感測器未能觸及之區域及人力監測通報，以達全面性並適用於城市應用。本研究所建立之模式包含兩個模組:圖像增強及影像辨識模組。首先，為改善監視器資料畫質普遍不佳及鏡頭固定使畫面呈單一性之缺點，本研究以Super Resolution General Adversarial Network (SRGAN)作為新穎影像增強作法來提升影像品質。接著，影像辨識模式模組使用Mask R-CNN及DeepLabv3+為基礎分別建構。最後，研究以實際淹水事件作為模式應用，以評估所開發之模式性能，結果證明是以DeepLabv3+模式結合影像增強作法表現為最佳，可以有效識別淹水發生及其範圍，並保有高效檢測效率可有效應用。在測試場次中，針對淹水辨識精確率為92%、召回率為94%，而用於評鑑辨識及實際重疊度的交併比(IoU)指標為87%；與水利署模式相比，本模式對於淹水範圍之判斷更為準確；針對模式適用性而言，於國、內外之影像畫面也能有效識別水體及應用。 此外，本研究亦證明預訓練權重及訓練數據對於模式有顯著影響，其中又以預訓練權重影響甚大，本研究所提出之影像增強作法，可有效提升模式整體性能，且模式辨識類別之範圍及準確度皆有所提升。總而言之，為彰顯淹水災害時快速且有效應變之重要性，本研究所提出之模式為提升城市的防災應變能力提供有效且準確之資訊，大幅降低淹水所造成生命財產損失之風險，且有助於提升城市的防災應變能力，以應對氣候變遷所帶來的威脅。

The frequency and intensity of short-duration heavy rainfall have escalated due to global climate change, increasing the probability of recurrent flooding and presenting a significant challenge to urban and underground water infrastructure. Information on the location of flooding zones and boundaries is crucial for disaster prevention and management decisions. Traditionally, flooding events are manually reported or are determined by flood sensors installed in critical areas of roads; however, these methods lack timeliness, incur installation and maintenance costs, and provide only regional monitoring data. Therefore, an effective and timely flood detection method is necessary. This study develops a deep learning model based on real-time flooding image recognition to replace flooding sensors and human monitoring using the existing widely distributed closed-circuit television (CCTV), ensuring comprehensiveness and the applicability of this method in urban applications. Images from intersection monitors are employed as data input. The model comprises two modules: image enhancement and image recognition. First, the super-resolution general adversarial network (SRGAN) is used as a image enhancement method to improve image quality by improving the pixel quality of the CCTV data and the uniqueness of the lens fixation. Next, the image recognition module is constructed using the mask region-based convolutional neural network (Mask R-CNN) and DeepLabv3+. This study addresses whether to employ pre-trained weights in modeling and the effect of training data on the model to analyze and determine the optimal model. Finally, this study utilizes an actual flooding event to evaluate the performance of the developed model. The results indicate that the model, particularly when using the DeepLabv3+ model combined with image enhancement techniques, achieves optimal performance. It effectively identifies the occurrence and extent of flooding while maintaining high detection efficiency. During the testing set, the model achieved a precision rate of 92% and a recall rate of 94% for flood identification, with an Intersection over Union (IoU) score of 87%, reflecting the overlap between predicted and actual flood extents. Compared to models used by the Water Resources Agency (WRA), this model demonstrates superior accuracy in determining flood boundaries. Furthermore, the model has been shown to be applicable in identifying water bodies and flood extents in both domestic and international image datasets. In conclusion, the proposed model offers significant potential for enhancing disaster prevention and management capabilities in urban areas, enabling more effective responses to flooding events driven by climate change.