Speaker
Description
Climate change is recognised as a major driver of extreme hydro-meteorological events worldwide. Rising temperatures, altered precipitation patterns, and extreme rainfall events have increased the frequency and intensity of floods, particularly in vulnerable regions. These challenges have encouraged the emergence of artificial intelligence (AI) as a cutting-edge technology for environmental monitoring, data analysis, and decision support. Recent studies have highlighted the crucial role of AI in addressing climate-related challenges through improved vulnerability assessment, predictive modelling, climate risk analysis, and early warning systems. By processing large volumes of climate, meteorological, satellite, and geospatial data, AI has enhanced predictive accuracy by 15–25%, particularly for climate forecasting and extreme weather event prediction. These capabilities have confirmed AI's potential for climate change adaptation, territorial resilience, and climate-related hazards, including floods.
Floods remain among the most frequent and damaging natural disasters, and artificial intelligence has demonstrated significant progress in mapping areas at risk of flooding and understanding both pre- and post-disaster conditions, thereby enhancing flood susceptibility mapping, flood forecasting, and early warning systems. Among AI approaches applied to flood management, machine learning (ML) and deep learning (DL) are the most widely used methods. ML algorithms, such as Random Forest (RF), Support Vector Machines (SVM), Logistic Regression, and XGBoost, are primarily used for flood susceptibility mapping and risk classification. Their advantages include their interpretability, low computational requirements, and their ability to analyse the relationships between multiple input data effectively. Comparative studies report high predictive performance, with ensemble learning approaches often achieving area under the curve (AUC) values ranging from 80% to 90%, and in some cases approaching 95%, using Random Forest (RF). In contrast, deep learning approaches are increasingly used for advanced prediction and forecasting tasks. Models such as artificial neural networks (ANN), convolutional neural networks (CNN), and long short-term memory (LSTM) networks can process large-scale, high-dimensional datasets and capture complex spatio-temporal dependencies, which makes them particularly suitable for real-time flood prediction in dynamic climatic conditions. More recently, hybrid models combining machine learning, deep learning, and conventional hydrological approaches have shown additional improvements in predictive accuracy (11%- 50% higher in accuracy compared to individual models), highlighting the advantages of combining different modelling techniques for flood risk management and decision support.
Although these studies demonstrate strong predictive capabilities, high predictive accuracy alone does not guarantee model robustness, transferability, or operational applicability in different geographical and climatic settings. Future directions address multiple needs, such as involving solutions that are both scalable and adaptable to various geographic and socio-economic conditions, emphasizing the critical need for more explainable AI systems to better inform stakeholders' decisions. Additionally, exploring novel AI models, particularly hybrid models and participatory approaches, will contribute to enhancing flood resilience.
Primary investigations indicate that current research focuses on improving the predictive performance of AI models, while little attention has been given to the selection of the study basin, despite its critical influence on data quality, model reliability, and research outcomes. To address these limitations, a decision support framework is designed to facilitate the selection of basins that utilize a multicriteria approach, considering multiple dimensions based on the literature review, such as Research Issues, Literature Gaps, Predefined Criteria, and Innovation Potentials.
The proposed framework was applied to four candidate sub-basins within the Tensift watershed in Morocco: Ourika, Zat, Rherhaya, and N'fis, for which detailed fact sheets were established to support a comparative analysis which considered a set of criteria derived from the proposed framework, including Physical and Hydrological Characteristics, Data Availability and Accessibility, Monitoring Network Density, Socio-economic Exposure, Research and Validation Potential, and Historical Flood Events.
The comparative application revealed significant differences among the candidate basins using the proposed framework. The Ourika basin was identified as the most suitable study area due to its flood history (Deadly floods, such as the one in 1995), dense hydrometric monitoring network, high availability of hydrological data(>40 years), and high socio-economic exposure(1644 economic establishments), Innovation and Research Gap Potential, which makes it a particularly relevant case study for future AI-based flood susceptibility modelling in semi-arid mountain environments. The results demonstrate the usefulness of combining literature-based knowledge with a decision-support process for study area selection, gathering the largest multiple set of criteria. Beyond the specific case of the Tensift watershed, the proposed framework may contribute to improving the reproducibility and operational relevance of AI-based flood studies in other climate-sensitive regions. The approach also supports broader climate adaptation objectives by facilitating the selection of representative study areas where AI models can be developed and transferred more effectively.
Future research will investigate hydro-climatic changes between drought and wetter periods within the selected basin to assess their impact on flood occurrence. The findings will support the comparison of AI-based flood susceptibility models and the exploration of innovative mitigation strategies for climate-resilient flood management.