影像辨識技術應用於河川堰塞湖形成之研究

Abstract

台灣位於環太平洋地震帶上，為地震活躍地區，每年平均有1000個有感地震，地震會造成土石鬆動，且近年來受氣候變遷的影響，極端氣候加劇，暴雨頻率和強度都有上升的趨勢，暴雨和地震容易誘發山坡地崩塌，造成大量的土砂堆積於河谷，進而導致堰塞湖的產生。堰塞湖是由於河道中的堵塞物而形成的湖泊，當堰塞湖潰決時，大量的洪水伴隨土砂下移，將對下游地區造成嚴重威脅，因此，若能夠快速且準確的辨識堰塞湖，對於水土防災乃為重要之課題。 有鑑於此，本研究提出了一種基於深度學習的自動化檢測堰塞湖之方法，以本研究區域雲林縣古坑鄉車心崙地區之目標河川2018年至2022年之閉路監視器 (Closed-Circuit Television, CCTV) 影像資料和網路所蒐集的水面影像建置數據集並標注，用以訓練實例分割模型¬—Mask R-CNN (Mask Region-based Convolutional Neural Networks) 來捕捉水面特徵，本研究提出之模型對水面辨識具有相當之準確性。測試的結果準確度 (Accuracy) 達到0.988，且於本研究提出之河道內準確度 (AccuracyR) 達到0.980，此外在精確率 (Precision) 、召回率 (Recall) 和F1-score則是有0.802、0.807和0.795之表現，能準確地辨識水面。 為自動判釋堰塞湖，本研究於Mask R-CNN水面辨識模型後銜接後處理模組，透過計算在框定之河道範圍內的上下游水面網格數，檢測水體範圍是否產生大幅度的變化，搭配制定出的門檻值與判斷條件，以此為根據去判定有無堰塞湖的產生，為解決本研究區域沒有歷史發生之堰塞湖影像資料，無法對模組進行驗證之問題，本研究繪製了四種可能發生之堰塞湖情境，分別為全部河道崩塌、上游河道崩塌、下游河道崩塌以及鏡頭外下游河道崩塌，使用以上四種堰塞湖模擬情境做為驗證模組能力之依據，結果顯示本研究提出之堰塞湖自動判釋模組具有很高的通用性與準確性，能夠精準的辨識出四種不同情境之堰塞湖。 本研究提出的基於 Mask R-CNN 的河川堰塞湖檢測模組具有很高的應用價值。該模式可以自動化地識別和判定河川中是否形成堰塞湖，不僅可以節省人力成本，更可即時通報相關權責機關做查證，並給予緊急調查及監測作為參考依據。

Taiwan is located on the Pacific Ring of Fire, a seismically active region, experiencing an average of 1,000 felt earthquakes annually. Earthquakes can cause soil loosening, and in recent years, climate change has intensified extreme weather, increasing the frequency and intensity of heavy rainfall. Both heavy rainfall and earthquakes can trigger landslides on slopes, resulting in large amounts of soil and sand accumulating in rivers, leading to the formation of Landslide-dammed lakes. A Landslide-dammed lake is a lake formed by obstacles in a river. When a Landslide-dammed lake collapses, a large scale of floodwater and sediment moves downstream, posing a significant threat to downstream areas. Therefore, the ability to quickly and accurately identify Landslide-dammed lakes is crucial for disaster prevention. In view of this, this study proposes an automated detection method for Landslide-dammed lakes based on deep learning. Using water surface images collected from the internet and closed-circuit television (CCTV) images of the target river area of Chexinlun from 2018 to 2022, a dataset was constructed and annotated to train an instance segmentation model—Mask R-CNN (Mask Region-based Convolutional Neural Networks) to capture water surface features. The proposed model in this study demonstrates considerable accuracy in water surface recognition. The test results show an accuracy of 0.988, and an accuracy within the river channel defined in this study (AccuracyR) of 0.980. Additionally, the model achieved a precision of 0.802, a recall of 0.807, and an F1-score of 0.795, effectively recognizing the water surface. To automatically identify Landslide-dammed lakes, this study integrate a post-processing module with the Mask R-CNN water surface recognition model. By calculating the number of water surface pixels upstream and downstream within the defined river channel, the changes of water surface are detected. Thresholds and conditional statements are defined to determine the formation of the Landslide-dammed lake. Due to the lack of historical Landslide-dammed lake images in the study area, which prevents model validation, four potential Landslide-dammed lake scenarios were drawn: total river channel collapse, upstream river channel collapse, downstream river channel collapse, and downstream river channel collapse outside the camera view. These four simulated Landslide-dammed lake scenarios were used to validate the model's capabilities. The results indicate that the proposed automatic detection model for Landslide-dammed lakes has high Generalization and accuracy, precisely identifying Landslide-dammed lakes in all four different scenarios. The Mask R-CNN-based river Landslide-dammed lake detection model proposed in this study has high application value. The model can automatically identify and determine whether a Landslide-dammed lake has formed in a river, saving labor costs and providing timely reports to relevant authorities for verification, emergency investigation, and monitoring reference.