Enhanced precipitation estimation in a Himalayan river basin through the fusion of multi-source datasets using various machine learning techniques

Harshita Tiwari

doi:10.1016/j.pce.2026.104418

Enhanced precipitation estimation in a Himalayan river basin through the fusion of multi-source datasets using various machine learning techniques

2026-05-23

0

PRE

AI

H

Harshita Tiwari *

A

Aditya Kumar Thakur

R

Rahul Dev Garg

DOI：10.1016/j.pce.2026.104418

Original

Original request for help

Save

Abstract

En 中文

Reliable precipitation estimation in mountainous basins is constrained by a sparse gauge network and complex topographic effects. While various open-access gridded precipitation products (GPPs) offer good temporal and spatial coverage, their utility at local and regional scales remains questionable. This may be attributed to relying on single GPPs or uniform models rather than location-specific, integrated approaches. In this study, we have assessed the individual performance of each GPP-model combination and present a Spatially Weighted Grid-Wise Ensemble Learning framework for the Budhi Gandaki basin in the central Himalaya. This study develops an integrated framework that systematically combines nine gridded precipitation products and observations from six rain gauges with four machine learning algorithms using grid-wise evaluation and adaptive station-weighting, producing a merged gridded precipitation product by blending the best-performing GPP-model pair available at each station through spatially weighted integration. Results indicate that some GPP-model pairs, like Asian Precipitation-Highly-Resolved Observational Data Integration Toward Evaluation (APHRODITE), when coupled with Random Forest (Root Mean Squared Error (RMSE) and Mean Squared Error (MSE) equals 5.09 mm/day at station S-1) and XGBoost (RMSE equals 3.94 mm/day at station S-2), showed superior performance in comparison to the other GPPs (RMSE range 8.61-10.14 mm/day). Conversely, satellite, reanalysis, and ensemblebased precipitation products dominate with increasing elevation. The integrated model effectively addresses spatial and performance gaps across individual GPP-model combinations, as evidenced by error propagation analysis at both model and station levels, which yielded a mean RMSE of 4.00 mm/day (ranging from 2.87 to 4.80 mm/day) and a positively skewed error distribution. The merged precipitation product showed colossal improvement in correlation coefficient (CC), rising up to 0.68 in comparison to GPPs, which had CC in the range 0.18 to 0.36. Overall, the study presents a scalable and adaptable framework for enhancing precipitation estimation in complex terrains, with demonstrated robustness across varying elevations and data conditions.

Keywords:

Integrated precipitation modeling

Grid-wise ensemble learning

Elevation-aware rainfall estimation

Multi-source data fusion

Adaptive spatial weighting

Enhanced precipitation estimation in a Himalayan river basin through the fusion of multi-source datasets using various machine learning techniques

Abstract

Journal

Organization

Cited Papers

Citing Papers