即時影像萃取土石流表面流速

Abstract

台灣山區地勢陡峭，河道坡陡流急，極易在強降雨事件中引發土石流災害。由於土石流具突發性與高速流動等特性，對鄰近居民與基礎設施造成極大威脅，因此即時監測與流速偵測技術之發展，為提升災害預警效能之關鍵課題。 本研究提出一套以即時影像為基礎之土石流表面流速萃取方法，利用影像中兩個感興趣區域（Region of Interest, ROI）之平均灰階值與其時間變化特徵作為分析依據。透過影像灰階化、平均灰階值平滑處理與斜率計算，建構兩組時間序列，並進行時間延遲分析。進一步結合均方根誤差（Root Mean Square Deviation, RMSD）進行最佳時間平移量判定，據以估算事件於兩ROI間之平均表面流速。此外，本研究引入浮動式門檻值以排除環境雜訊干擾，並輔助判斷事件進出ROI之時段。 本研究分別透過人造數值影像、室內水槽實驗與現地監測影像三種方式進行驗證。數值實驗結果顯示，在理想條件下，本方法可準確估算流速，誤差約小於1.4%。室內水槽實驗中，由於光源變化與相機ISO及光圈設定為自動模式，導致影像亮度不一致，產生約-20.22%的誤差。至於現地土石流影像，受限於天候因素與設備限制，影像多處於模糊或沾附水滴、霧氣等情形，雖部分案例誤差可低至15.13%，惟整體誤差偏高，亦存在多起誤判狀況。 整體而言，本研究所提出之方法可應用於土石流流速之初步推估，但於現地實務應用上，影像品質受環境與氣候條件限制，仍為其準確性之主要影響因素，後續尚需針對此部分進行改善與強化。

Taiwan's mountainous terrain and rapid streams make the region prone to debris flows during heavy rainfall, posing serious risks to people and infrastructure. Real-time monitoring and velocity detection are crucial for early warning. This study presents an image-based method to extract debris flow surface velocity in real time. Two Regions of Interest (ROIs) are analyzed by tracking average grayscale values and their changes over time. Using smoothed grayscale time series and slope data, a time delay analysis with Root Mean Square Deviation (RMSD) determines the optimal shift to estimate surface velocity. A floating threshold mechanism filters environmental noise and detects event timing within ROIs. This study was validated using synthetic images, indoor flume experiments, and field debris flow video. In ideal conditions, the method showed high accuracy with an error below 1.4%. In flume experiments, lighting variation and automatic camera settings caused brightness inconsistency, leading to an error of about -20.22%. Field images were affected by poor visibility, water droplets, and mist, resulting in higher errors, though some cases achieved a minimum error of 15.13%. Overall, the proposed method is effective for preliminary surface velocity estimation. However, its accuracy in field applications is limited by environmental and weather-related image quality issues, highlighting the need for further improvements.