土石流監測區域自動判識

Abstract

近年來，攝影機已廣泛設置於土石流觀測站，但因視角與架設位置受限，河道在影像中經常僅占部分區域，加上植物搖動與暴雨造成的灰階值劇烈變動，易降低土石流相關參數的偵測準確度。為減少背景干擾並提升運算效率，本研究提出一套自動化河道區域（ROI）提取流程，使影像分析專注於實際河道位置。 本研究蒐集不同光照、天氣與地點的影像，分別計算灰階值、平均時變量與空間變異度三種特徵，並以 K-means 進行三值化，形成 27 種特徵組合。接著透過統計分析歸納出代表水體與溪床（含沙岸）的組合，再利用形態學操作，萃取出主要水體及其鄰近岸邊。 ROI 之準確性以兩項客觀標準驗證：無事件期間以五組人工標註結果評估重疊率；事件期間以事件最大覆蓋範圍計算功能性準確率。結果顯示，本文流程生成之 ROI 不僅在事件前能提取合理的河道位置，事件發生時亦能涵蓋土石流的主要流動範圍。 最後透過現地事件水位超出原始 ROI 的情形率定擴張比例，建立統一的 ROI 擴大標準，使後續案例能依該標準自動調整監測範圍。

In recent years, cameras have been widely deployed at debris-flow monitoring stations; however, due to limitations in viewing angles and installation positions, the river channel often occupies only a small portion of the image. In addition, background disturbances—such as vegetation movement or heavy rainfall—cause significant fluctuations in pixel intensities, reducing the accuracy of debris-flow parameter detection. To mitigate these effects and improve computational efficiency, this study proposes an automated river region of interest (ROI) extraction procedure that focuses image analysis on the actual river area. Images collected under various lighting, weather, and site conditions were used to compute three pixel features: grayscale intensity, temporal variation, and spatial variability. Each feature was classified into three levels using K-means clustering, forming 27 feature combinations. Statistical analysis was then applied to identify combinations most representative of water and rocky areas (including sand and gravel). Morphological operations were further used to extract the main river water region and adjacent riverbanks. The accuracy of the extracted ROI was evaluated using two objective criteria: (1) overlap with five sets of manually annotated river boundaries during non-event periods, and (2) functional accuracy based on the maximum event coverage during debris-flow events. Results show that the proposed method can reliably extract meaningful river regions during normal monitoring conditions and effectively cover the main flow area during debris-flow events. Finally, by analyzing the extent to which actual event water levels exceeded the initial ROI, this study establishes a unified ROI expansion standard. Subsequent cases can automatically adjust their monitoring range according to this calibrated expansion ratio.