應用HRNet模型建立自動水位量測系統

Abstract

隨著氣候變遷導致極端降雨頻率增加，河川水位監測在水利防災中扮演愈加關鍵的角色。然而，傳統水位監測方式多依賴實體水尺與人工判讀，不僅在偏遠地區佈建困難，也容易受環境因素干擾而降低辨識準確度。為解決此問題，本研究提出一套結合電腦視覺與虛擬水尺的自動水位量測方法，透過深度學習模型自動辨識影像中的水面線，並轉換為實際水位高度資訊。 本研究採用 High-Resolution Network (HRNET)架構進行水面線位置的回歸訓練，設計並輸出 20 個等間距的垂直坐標。訓練資料來源涵蓋實地攝影與實驗水槽影像，並搭配人工標註進行模型訓練。為提升模型辨識能力，本研究設計資料擴增策略，包括外觀變化擴增（ADA）與方向性隨機擴展（RED），並透過困難樣本補強與劣質樣本剔除來優化訓練資料集，解決模型訓練錯誤與誤判問題。預測結果經由影像正射校正與虛擬水尺的轉換，可準確換算成水位高程。後處理方面，則引入多張影像時間平均與傅立葉分析，提升輸出穩定性與應用可靠度。 實驗結果顯示，最終模型的平均絕對誤差（MAE）為 0.0157，決定係數（R²）達 0.9874，在 2% 誤差容忍度下準確率為 84.97%，10% 誤差容忍度下準確率達 99.5%。本研究所建構系統在多樣化的影像情境下具備良好的水面線辨識與水位估算能力，並於特定誤差範圍內達成高準確率。本研究不僅展示深度學習模型應用於無水尺環境下水位判讀之可行性，也證明透過樣本資料優化與後處理設計，能有效提升系統於實務應用中的穩定性，為智慧型水文監測提供新的技術方向。

With the increasing frequency of extreme rainfall events driven by climate change, river water level monitoring has become a critical component of disaster prevention and water resource management. However, conventional methods relying on physical staff gauges and manual interpretation face challenges in installation, accuracy, and efficiency, especially in remote or harsh environments. To address these limitations, this study proposes an automated water level measurement system integrating computer vision and virtual gauge technology. The system employs a deep learning model based on the High-Resolution Network (HRNet) architecture to detect the water surface line in images and convert the output into actual water level estimations. The model performs a regression task by predicting the vertical coordinates of 20 equally spaced points along the water surface. The model is trained on images from field sites and laboratory flumes, with manually labeled ground truth. To enhance model robustness, data augmentation strategies, including Appearance-based Data Augmentation (ADA) and Random Extension in Direction (RED), are introduced along with difficult sample supplementation and poor-quality sample removal for dataset optimization. Predicted results are orthorectified and mapped to virtual gauge elevations, and further stabilized using multi-frame averaging and Fourier-based wave analysis. Experimental results demonstrate that the proposed system achieves exceptional accuracy with a Mean Absolute Error (MAE) of 0.0157, coefficient of determination (R²) of 0.9874, accuracy rates of 84.97% at 2% error tolerance and 99.5% at 10% error tolerance. The system performs reliably under diverse environmental conditions without requiring physical water gauges. This study validates the feasibility of applying deep learning to automated water level estimation in natural environments and highlights the value of dataset optimization and post-processing in enhancing reliability and field applicability. The research provides a novel technical approach for intelligent hydrological monitoring systems.