Water security monitoring and forecasting using artificial intelligence: application of artificial neural networks in hydrological ecosystem services modeling

• Ano de defesa: 2024
• Autor(a) principal: Leon Sarmiento, Jorge Eduardo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Artificial intelligence; Hydrological modeling; Inteligência artificial; Modelagem hidrológica; Remote sensing; Sensoriamento remoto
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/91/91131/tde-10102024-105155/

This research work explores the use of artificial neural networks (ANN) in hydrological modeling. The study is based in 4 different experiments to test the ability of ANN to solve common hydrology modeling tasks usually addressed by the use of rational or empirical modeling. These 4 experiments start under control situations, where hydrological data generated by hydrological and water-energy balance models are used to test the capacity of ANN to infer or learn the concepts behind our own understanding and modeling capacities of these processes, and increase in complexity by replacing control data with real world data where inexact, incomplete and multi-modal (field instrumentation measurements, satellite imagery capture) data is found. In this study the ANN were evaluated also in terms of their own architecture by using 3 of the most common type of neural networks (NN) usually associated with tabular and time series data analysis: MLP-Feedforward NN, LSTM-Recurrent NN and Transformers based NN. While the trend at the moment of writing this document regarding earth systems modeling, which certainly include the hydrology topic of this research, is to favor high performance and cloud computing approaches, all the experiments in this research were designed and executed using consumer grade equipment and free access computing resources. The results of these experiments demonstrated the common hydrological modeling tasks including water balance modeling and forecast, usually addressed by the use of dynamic physically-based (semi)distributed models, can be accomplished by the use of ANN outperforming model performance metrics (NSE index) in published results of previous USP dissertations, and considerably reducing the amount of time required to produce results. However, the most interesting aspect of using ANN in hydrological modeling is the capability to model processes or time frames for which few or no hydrological models are available, or by using biophysical data available from remote sensing that is incompatible with rational models commonly used. Finally, while this research explores the potential to use these technologies in addressing common hydrological modeling task, and in some cases demonstrate a significant advantage on using these tools instead of traditional hydrological models, the fundamental question of how can this improve water security is only partially addressed because improving our water resources management is not only a technical challenge. While is true that having faster and/or more precise forecast of hydrological processes like peak flows, floods or droughts by using ANN can suppose an improvement regarding information access, any water security improvement as a whole depends in the end on the capacity of governance institutions to use this and other available scientific insights on their decision making process.