Parsimonious Neural Prediction of Bank Erosion and Sediment Deposition in a Tropical Distributary Channel
IOSR Journal of Mechanical and Civil Engineering · 2026-05 · 23 (5) · 32-43
DOI: 10.9790/1684-2305033243 ↗Kurz gefasst
Fachartikel in der Zeitschrift IOSR Journal of Mechanical and Civil Engineering, 2026-05. Verfasst von Anthony Enifome Eseha und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Bank erosion and sediment deposition along tropical deltaic distributaries threaten land, infrastructure and navigation, yet physically based bank-stability models are difficult to parameterize where field data are sparse. This paper reports a deliberately compact artificial neural network that predicts bank erosion and sediment deposition simultaneously along a 5,000 m reach of the Erowha River, a distributary of the Forcados system in Delta State, Nigeria. A total of 101 chainage-referenced stations (L1–L101) at 50 m intervals were surveyed during the wet season, yielding paired measurements of channel geometry, bank slope, velocity, discharge, angle of internal friction, cohesion, land use and vegetation cover alongside the two response variables. Seven predictors: stream depth, top width, bottom width, bank slope, velocity, discharge and angle of internal friction formed the network inputs. Ten multilayer-perceptron topologies were trained by Levenberg–Marquardt backpropagation with a 70:15:15 training–validation–testing partition, with normalisation parameters fitted within each training fold to prevent information leakage. The selected 7–1–2 network, with only 12 trainable parameters, achieved a weighted coefficient of determination of 0.9553, an erosion R² of 0.9794 (MAPE 1.385%) and a deposition R² of 0.7960 (MAPE 2.832%). Measured erosion (7.90–12.10 kg/m²) correlated almost perfectly with discharge (r = 1.000) and velocity (r = 0.996), while vegetation cover (45.00–65.00%) was perfectly inversely associated with erosion (r = −1.000). Larger networks degraded sharply: a 7–11–2 topology collapsed to a deposition R² of −3.8887. At this sample size, parsimony rather than capacity governs generalisation.
Zitieren
APA
Eseha, A. E., Nwoke, H. U., & Dike, B. U. (2026). Parsimonious Neural Prediction of Bank Erosion and Sediment Deposition in a Tropical Distributary Channel. IOSR Journal of Mechanical and Civil Engineering, 23(5), 32-43. https://doi.org/10.9790/1684-2305033243
Vancouver
Eseha AE, Nwoke HU, Dike BU. Parsimonious Neural Prediction of Bank Erosion and Sediment Deposition in a Tropical Distributary Channel. IOSR Journal of Mechanical and Civil Engineering. 2026;23(5):32-43. doi:10.9790/1684-2305033243
BibTeX
@article{Eseha2026,
author = {Eseha, Anthony Enifome and Nwoke, Herbert Ugwueze and Dike, Benedicta Uchenna},
title = {Parsimonious Neural Prediction of Bank Erosion and Sediment Deposition in a Tropical Distributary Channel},
journal = {IOSR Journal of Mechanical and Civil Engineering},
year = {2026},
volume = {23},
number = {5},
pages = {32--43},
doi = {10.9790/1684-2305033243}
}
EUVIMED