Neural classification under discontinuous hydrometeorological data for flash flood early warning
Keywords:
data-driven hydrological modeling, neural networks, water level forecasting, flash flood early warningAbstract
This study presents the development of a data-driven hydrological model for one-hour-ahead water-level forecasting using a Multi-Layer Perceptron (MLP) neural network. The model was trained with water level and rainfall data from 12 monitoring stations located in a small, fast-responding watershed prone to flash floods. Training data span the period from 2020 to 2023, while an independent dataset from 2024 was used for testing. The MLP was initially implemented as a regression model, but its performance was limited by gaps in historical rainfall records. To overcome this limitation, the problem was reformulated as a classification task, in which forecasted water levels were grouped into three hydrologically meaningful classes representing non-critical, intermediate, and alert conditions. A sequence of classification experiments was conducted to refine the model, including the explicit treatment of missing rainfall data and an alternative class-definition strategy incorporating antecedent water levels and rainfall accumulation to represent pre-alert conditions. The proposed classifier achieved stable performance and improved consistency in identifying alert situations, providing a robust and operationally meaningful framework for short-term flood early-warning systems in data-limited environments.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Revista de Gestão de Água da América Latina (Water Management Journal in Latin America)
This work is licensed under a Creative Commons Attribution 4.0 International License.
When sending the manuscript, notice that:
To be responsible for the other authors, when applicable, as co-responsible for the technical and scientific content of the article according to Article 5 of Brazilian Law N. 9610, regarding Copyright.
All statements published in the manuscript are the sole responsibility of the authors. However, all published material becomes the property of REGA, which reserves the copyright. Therefore, no material published in REGA may be reproduced without the written permission of REGA. All authors of articles submitted to REGA must sign a Copyright Transfer Agreement, which will take effect from the date of acceptance of the article. This term will be requested by REGA prior to the publication of the article. The author responsible for the article will receive, free of charge, the electronic record of the publication (in PDF format).
All articles published open access will be immediately and permanently free for everyone to read, download, copy and distribute. Permitted reuse is defined by your choice of one of the following user licenses:
Creative Commons Attribution (CC BY): lets others distribute and copy the article, to create extracts, abstracts, and other revised versions, adaptations or derivative works of or from an article (such as a translation), to include in a collective work (such as an anthology), to text or data mine the article, even for commercial purposes, as long as they credit the author(s), do not represent the author as endorsing their adaptation of the article, and do not modify the article in such a way as to damage the author’s honor or reputation.
Author Rights
For open access publishing this journal uses an exclusive licensing agreement. Authors will retain copyright alongside scholarly usage rights and REGA will be granted publishing and distribution rights.
Author Self-Archiving Policy
This journal permits and encourages authors to post items submitted to the journal on personal websites and institutional or funder repositories after publication. The final published PDF version should be used and bibliographic details that credit the publication in this journal should be included.
