Inter-comparison and mechanistic interpretation of deep learning models for turbidity prediction in rural areas

Freshwater ecosystems are closely related to human lives, especially in rural areas, necessitating advanced tools for monitoring sediment dynamics, which are critical for water quality and aquatic health. Turbidity, as one of the key indicators of suspended sediments, is widely adopted to monitor the sediment dynamics. This paper addresses the urgent need for robust turbidity prediction in Hong Kong's Lai Chi Wo (LCW) catchment, as a case study where rural communities rely on riverine water resources. Four DL models namely Long Short-Term Memory (LSTM), Encoder-Decoder LSTM (ED-LSTM), Temporal Convolutional Network (TCN), and Informer are employed for multi-step turbidity forecasting using high-resolution (5-minute interval) on-site hydrological data. Our results demonstrate Informer's superior short-term (ST) prediction skill (NSE = 0.83, KGE = 0.85), closely followed by ED-LSTM (NSE = 0.80, KGE = 0.80). Both models exhibit temporal dependency modeling, outperforming LSTM and TCN, particularly in long-term (LT) forecasts. Moreover, the insight of temporal modeling is uncovered by element-wise saliency maps, a model explainable tool, revealing turbidity autoregression and rainfall triggers as dominant drivers. The merit of predictions of ED-LSTM and Informer originates from their capacity in capturing fading temporal dependencies and identifying the physical trigger in models. This work advances DL applications in hydrology by bridging the gap between prediction performance and mechanistic interpretation, offering new insights for predictive modeling.