GeoAI tech offers smarter, faster and more transparent flood predictions

To prevent data imbalance, a common problem in environmental modeling where non-flooded areas vastly outnumber flooded points, the researchers employed random under-sampling techniques, ensuring the algorithms remained unbiased and sensitive to minority classes (actual flood events). The models were trained on 70% of the data and tested on the remaining 30%.

Researchers from Istanbul Technical University have unveiled a cutting-edge framework that could reshape how governments and urban planners anticipate and mitigate flood disasters. By merging geospatial artificial intelligence (GeoAI) with explainable AI (XAI), the researchers have designed a transparent and highly accurate system capable of pinpointing flood-prone zones across Türkiye’s most densely populated and industrially significant region, Marmara.

Published in Systems, the study “Flood Risk Analysis with Explainable Geospatial Artificial Intelligence (GeoAI) Techniques” introduces a hybrid data-driven model that integrates machine learning algorithms with classical hazard analysis to forecast flooding with unprecedented precision. The framework not only identifies high-risk areas but also reveals the most influential factors behind flood susceptibility, setting a new standard for disaster management in rapidly urbanizing regions.

Why floods demand a new approach to disaster risk management

The study explores how escalating flood frequency and severity across Türkiye are driven by the combined forces of climate change, erratic precipitation, and uncontrolled urban expansion. Traditional flood management methods, often based on historical data or static models, have proven insufficient in the face of today’s rapidly shifting environmental and socioeconomic dynamics. The Marmara Region, which encompasses eleven provinces including Istanbul, Bursa, Kocaeli, and Çanakkale, represents a critical case: home to more than 26 million people and serving as the country’s industrial heartland, it suffers from both geographic vulnerability and infrastructural strain.

Between 1975 and 2012, Türkiye recorded 889 flood events, resulting in hundreds of deaths and substantial material losses. The Marmara Region alone experienced hundreds of these incidents, including the record-breaking rainfall and widespread flooding of September 2009. Such disasters revealed the weaknesses of existing drainage systems and the urgency for predictive, rather than reactive, flood management tools.

The researchers argue that effective disaster management must extend beyond reactive crisis response to proactive resilience planning. The proposed GeoAI framework merges classical analytical methods such as the Analytical Hierarchy Process (AHP) with modern machine learning models, ensuring that both human expertise and data-driven precision are incorporated. This dual approach, they contend, not only enhances accuracy but also makes flood prediction explainable and actionable for policymakers.

The model evaluates multiple environmental and anthropogenic variables, including rainfall intensity, drainage density, proximity to waterways, population density, elevation, land use, and topography, to generate a regional flood risk profile. The integration of explainable AI enables transparency by showing how each factor contributes to the overall prediction. This transparency is essential for public trust and for aligning disaster policy with scientific reasoning.

How GeoAI and machine learning drive precision in flood forecasting

Using data collected from open and official sources such as the General Directorate of Meteorology (MGM), the State Hydraulic Works (DSI), and the US Geological Survey (USGS), the study trained and tested two leading machine learning algorithms, Random Forest (RF) and Extreme Gradient Boosting (XGBoost).

Each algorithm was used to classify flood-prone zones by analyzing 13 geospatial parameters, including topographic curvature, drainage proximity, precipitation trends, soil infiltration, land use, and population distribution. The datasets, covering a 50-year flood inventory and 40 years of precipitation data, were processed at a 30-meter resolution to ensure high spatial accuracy.

To prevent data imbalance, a common problem in environmental modeling where non-flooded areas vastly outnumber flooded points, the researchers employed random under-sampling techniques, ensuring the algorithms remained unbiased and sensitive to minority classes (actual flood events). The models were trained on 70% of the data and tested on the remaining 30%.

When evaluated using standard performance metrics such as precision, recall, F1 score, and the area under the ROC curve (AUC), XGBoost emerged as the superior model. It achieved an accuracy of 94.9%, precision of 91.1%, F1 score of 94.9%, and AUC value of 0.978, outperforming Random Forest across all categories. These results demonstrate that the XGBoost algorithm offers robust generalization capabilities and strong predictive power even in complex, heterogeneous landscapes.

The explainability component was powered by SHapley Additive exPlanations (SHAP), an XAI technique that interprets how each variable influences the final prediction. SHAP analysis revealed that drainage proximity, elevation, land use, and precipitation were the most influential determinants of flood vulnerability. Drainage proximity alone emerged as the dominant driver, underscoring the region’s infrastructural deficiencies and the critical role of hydraulic design in urban flood prevention.

By integrating AHP’s structured decision-making framework with XGBoost’s non-linear learning capability, the researchers created a hybrid model that balances expert judgment with data adaptability. The AHP provided the initial weights for hazard criteria, for example, waterways proximity (23%), waterbodies proximity (31%), and flow direction (19%), while the machine learning algorithm fine-tuned these values through data-driven optimization. This synergy achieved not only higher precision but also improved interpretability, allowing decision-makers to understand why certain locations are at greater risk.

Mapping the future of flood resilience through explainable AI

The Marmara Region flood risk map produced by the study revealed a clear geographical distribution of hazards and vulnerabilities. High and very high-risk clusters were concentrated in the coastal districts of Istanbul, the Çanakkale Strait, the southern shores of Balıkesir, and the urban centers of Kocaeli and Sakarya. These areas combine flat terrain, dense population, and inadequate drainage networks, making them especially prone to inundation during heavy rainfall.

On the other hand, interior provinces like Bilecik, Edirne, and parts of Bursa showed moderate to low risk due to their higher elevations and better water dispersion characteristics. The model’s predictive output showed a strong correlation with historical flood data, validating its accuracy. The hybrid framework successfully captured the spatial patterns of past flood events, including those linked to coastal surges and high precipitation zones.

Overall, machine learning and traditional methods should not be viewed as competitors but as complementary tools for building comprehensive risk assessment systems. The integration of Multi-Criteria Decision Making (MCDM) methods such as AHP with GeoAI allows for a balanced approach that combines the consistency and structure of expert-based analysis with the scalability and adaptability of artificial intelligence.

From a governance perspective, the findings emphasize the need to translate risk maps into policy actions. The researchers recommend several immediate applications, including the installation of automatic water-level monitoring stations, the deployment of SMS-based early warning systems, and the prioritization of urban drainage investments such as rainwater lines, retention basins, and infiltration pits. They also suggest that local governments update zoning and emergency response plans using the new flood risk maps to enforce construction restrictions in high-risk zones.

In addition, the authors highlight the potential for scaling this model beyond Marmara to other flood-prone regions globally. The data-driven, open-source framework makes replication feasible for local authorities with limited resources. However, they caution that the accuracy of the system depends on the quality and resolution of available data. The flood inventory data, often stored as point vectors rather than comprehensive hydrological layers, may limit model precision in some cases.

Despite such constraints, the proposed hybrid methodology demonstrates that combining GeoAI and XAI can achieve both predictive excellence and interpretive clarity, two characteristics rarely seen together in environmental modeling. The model’s performance improvements, including a 5% gain in AUC and 4% rise in overall accuracy over conventional rule-based methods, reflect the value of fusing human expertise with machine learning.

The broader implication of this research extends beyond hydrology. The study reinforces the role of explainable AI in public governance and climate resilience. Transparent AI-driven systems can bridge the trust gap between scientific models and policy implementation, ensuring that risk management strategies are not only effective but also publicly accountable.