Summer Research Internship · Mentor: Prof. Tianfang Xu

Channel Attention ConvLSTM for Streamflow Prediction in Snow-Dominated Watersheds

Rijul Dahiya¹

¹Arizona State University · School of Sustainable Engineering and the Built Environment

Abstract

Snowmelt plays an important role in supplying and sustaining streamflow in the arid and semi-arid western United States. Accurate streamflow prediction in snow-dominated karst watersheds is difficult due to spatially varying snow accumulation and melt and interconnected karst and porous groundwater systems. We develop a Channel Attention Module over a ConvLSTM model to simulate daily streamflow in the Logan River watershed, achieving NSE of 0.846 and PBIAS of 0.038% on testing data, an 8.33% improvement over prior ConvLSTM versions.

Key Contributions

Channel Attention Module integrated over ConvLSTM to adaptively weight spatially varying snow and groundwater signals.
Sequential spatial and channel attention mechanisms for refined feature representation.
8.33% improvement in streamflow quantification over the previous ConvLSTM version.
Evaluated on the Logan River watershed (581 km²) with 10-year test period (WY2009–WY2018).

Results

Evaluated over the WY2009–WY2018 testing period. The PBIAS of +0.038% indicates the model very slightly underestimates total streamflow.

Model	NSE ↑	PBIAS (%)
Channel + Spatial Attention ConvLSTM (sequential)	0.846	0.038

Figures

Figure 1. Logan River watershed (581 km²) located in the Bear River mountain range, study area for snowmelt-driven streamflow prediction.

Figure 2. Three-layer ConvLSTM architecture with Channel + Spatial Attention Module for adaptive feature refinement.

Figure 3. Predicted vs. true streamflow, model achieves NSE 0.846, closely tracking observed discharge patterns.

Figure 4. Model training loss graph, convergence of the Channel Attention ConvLSTM over training epochs.

Full paper