Logo image
Back
Contribution of Hydrological Feature Engineering and Multi-Source Data Fusion to Rainfall Imputation Accuracy: A Case Study in the Tweed River Catchment, Australia
Conference proceeding

Contribution of Hydrological Feature Engineering and Multi-Source Data Fusion to Rainfall Imputation Accuracy: A Case Study in the Tweed River Catchment, Australia

Banujan Kuhaneswaran, Golam Sorwar, Ali Reza Alaei and Feifei Tong
Moratuwa Engineering Research Conference (MERcon), pp.301-306
2026 Moratuwa Engineering Research Conference (MERCon) (Moratuwa, Sri Lanka, 13/08/2026–14/08/2026)
18/09/2026
DOI: https://doi.org/10.1109/MERCon71835.2026.11691536
Appears in  Recent Faculty of Science and Engineering Publications

Metrics

1 Record Views

Abstract

feature engineering forecasting gridded data imputation rainfall
Ground-based rainfall gauge networks in flood-prone catchments are characterised by substantial missing data, yet existing imputation methods evaluate performance solely through direct error metrics without examining operational consequences for forecasting. A further challenge is the circular dependency that arises when hydrological features are engineered from an incomplete series of data, as rolling statistics and antecedent condition indicators cannot be reliably computed over windows containing missing values. This paper presents a structured framework comprising two-stage feature engineering, a three-case ablation imputation experiment, and rainfall forecasting validation. The framework was evaluated across 13 Bureau of Meteorology stations in the Tweed River Catchment, New South Wales, Australia. Five imputation algorithms, namely Random Forest (RF), K-Nearest Neighbours (KNN), LSTM, SAITS, and BRITS, were tested under three progressively enriched input feature configurations: (i) temporal only, (ii) feature-enhanced, and (iii) feature-enhanced with blended SILO-IMERG gridded rainfall data fusion. RF achieved the strongest imputation performance, reducing mean RMSE from 18.34 mm to 5.12 mm under full integration, with per-station improvements of 49% to 84%. Feature engineering accounted for the majority of this gain, with gridded data fusion providing a consistent but smaller marginal contribution. Rainfall forecasting experiments using all 17 imputed datasets confirmed that imputation strategy selection has measurable consequences for forecasting accuracy, with RF-imputed inputs yielding the highest KGE of 0.756 and 0.702 under LSTM and GRU models, compared to 0.626 and 0.582, respectively, under the temporal-only baseline.

Details

Logo image