Improving precipitation estimation and hydrological simulation in Tianshan Mountain basins via CNN-SE-EF fusion

Study region: The Tianshan Mountains, a cold high-mountain, arid–humid transition zone with complex topography and mixed rain–snow runoff generation. We use 2000–2020 daily data from 19 stations as reference to evaluate and fuse multi-source precipitation for hydrologic application in the Tailan River basin and surrounding areas. Study focus: We benchmark six precipitation products (CHM, CMORPH, ERA5-Land, GPM IMERG, PERSIANN, TRMM) using continuous (R², MAE, RMSE, BIAS) and event metrics (POD, FAR, CSI). To address nonlinear spatiotemporal structure and leverage atmospheric controls, we design a CNN–SE–EF fusion that couples a convolutional backbone with squeeze-and-excitation attention and five covariates (2-m temperature, 2-m dew point, 10-m wind u/v, surface pressure). Regional transferability is tested via extended triple collocation (ETC); hydrologic utility is assessed by forcing the Snowmelt Runoff Model (SRM). New hydrological insights for the region: At low elevations, CHM and ERA5 perform best (lower FAR, higher POD/CSI, smaller MAE/RMSE, near-zero BIAS), whereas CMORPH and PERSIANN in high relief show higher false alarms and systematic underestimation. CNN–SE–EF outperforms CNN–SE and Bayesian averaging in R²/MAE/MSE, exhibits stronger cross-station stability, and delivers spatial skill superior to CLDAS v2.0 and GPCC. Fused precipitation improves SRM streamflow in the Tailan River (calibration/validation R² ≈ 0.76/0.91; volume bias ≈ 12.3 %), with remaining wet-season peak biases linked to simplified snow–ice and routing representations. The scheme is transferable to ungauged cold high-mountain basins.