Reconstruction of daily rainfall data using the concepts of networks: Accounting for spatial connections in neighborhood selection

Tiwari, Shubham; Kumar Jha, Sanjeev; Sivakumar, Bellie

Abstract

Accurate and reliable rainfall data is one of the fundamental prerequisites in hydrological modelling. The rainfall data at a desired location can be reconstructed using interpolation methods, such as Inverse Distance Weighting (IDW), which is frequently used in hydrology. In standard IDW neighbors are selected based on geographical proximity or nearest neighbor (IDW_NN). However, in a basin with variable topography, nearby rain gauges may be located at very different elevations and, thus, they may not accurately represent the spatial connection in rainfall. In this work, the theory of networks, with nodes and links as the basis, is applied to select neighbors while applying IDW. Two variants of neighbor selection models are proposed: IDW with linked neighbours (IDW_LN) and IDW with clustered neighbors (IDW_CN). For reconstruction, thirty years of daily rainfall data from 430 rain gauges in Murray-Darling Basin (MDB) are utilized. To evaluate the performance of the proposed models, one-station-leave-out cross validation approach is used and the associated Root-Mean-Squared-Error (RMSE) and Bias-percentage (BP) are calculated. Different values of number of neighbors (n), Correlation Threshold (CT) and Clustering Coefficient Range (CCR) are used to measure the errors associated with the proposed models. On comparing with IDW_NN, results show that reconstruction using IDW_LN has lower RMSE at about 30 percent of stations and lower BP for about 50 percent of stations; while IDW_CN shows lower RMSE at about 25 percent of stations and lower BP for about 45 percent of stations. The IDW_NN performed better than IDW_LN and IDW_CN at more than 50 percent of stations though the average error associated with all the three models are comparable for all CT values. In a natural system, a concept like traditional IDW (IDW_NN) may be more accurate than the network-based approach (IDW_LN and IDW_CN) but may not be completely efficient in accounting the spatial rainfall variability. The encouraging results for the reconstruction of rainfall in this study seem to indicate that the approach can be further helpful in the reconstruction of a wide range of meteorological parameters with spatial correlation.

Citation

Tiwari, S., Kumar Jha, S., & Sivakumar, B. (2019). Reconstruction of daily rainfall data using the concepts of networks: Accounting for spatial connections in neighborhood selection. Journal of Hydrology, 579, Article 124185. https://doi.org/10.1016/j.jhydrol.2019.124185

• Journal Article Type: Article
• Acceptance Date: Sep 26, 2019
• Online Publication Date: Oct 3, 2019
• Publication Date: 2019-12
• Deposit Date: Jun 15, 2020
• Journal: Journal of Hydrology
• Print ISSN: 0022-1694
• Publisher: Elsevier
• Peer Reviewed: Peer Reviewed
• Volume: 579
• Article Number: 124185
• DOI: https://doi.org/10.1016/j.jhydrol.2019.124185
• Public URL: https://durham-repository.worktribe.com/output/1268470