Document Type : Research Article
Authors
Department of Physical Geography, Faculty of Social Sciences, University of Mohaghegh Ardabili, Ardabil, Iran
Abstract
Flood risk management using machine learning methods for identifying susceptible areas plays an effective role in reducing flood damages. This study aimed to compare the performance of machine learning algorithms, including Random Forest (RF), Gradient Tree Boosting (GTB), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN), for flood susceptibility mapping in the Doostbiglo watershed using the Google Earth Engine cloud-based processing environment. A total of 585 flood and 699 non-flood points were extracted from Sentinel-1 imagery and used for model training and validation. The factors affecting flood occurrence, including elevation, slope, runoff, precipitation, Stream Power Index (SPI), Topographic Wetness Index (TWI), drainage and road density, Normalized Difference Vegetation Index (NDVI), soil texture, geological formations, and land use, were considered as model inputs. Flood susceptibility maps were prepared in five classes, ranging from very low to very high, and model performance was evaluated using the ROC curve, AUC, Kappa coefficient, and overall accuracy. The results showed that all four models mainly identified the eastern and northern areas and parts of the western slopes of Sabalan as having high and very high susceptibility. The GTB model, with an AUC of 0.965, showed the highest ability to discriminate flood-susceptible areas and classified 21.94% of the study area as having high and very high susceptibility. The RF, SVM, and KNN models, with AUC values of 0.964, 0.926, and 0.920, respectively, also showed acceptable performance. The map produced by the GTB model can be used to identify flood-susceptible areas and prioritize flood risk management and mitigation measures.
Introduction
Flood, one of the most destructive natural hazards, causes enormous damage to infrastructure and communities annually, and its effective management requires a thorough understanding of the factors contributing to its occurrence and the identification of high-risk areas. Despite technological advances, the lack of sufficient data and the complexity of flood patterns continue to pose challenges to the accurate assessment of flood-susceptible areas. Flood susceptibility mapping is considered an essential step toward preventing and reducing the adverse impacts of this phenomenon. Factors such as climate change, improper watershed management, deforestation, and overgrazing play a significant role in intensifying floods, particularly in Iran. Nowadays, advanced methods based on artificial intelligence and machine learning, with their ability to model complex relationships and provide high accuracy, have brought about major improvements in the accuracy of flood prediction and susceptibility mapping. The present study aimed to compare the performance of advanced machine learning algorithms in flood susceptibility mapping of the Doostbiglo watershed. The novelty of this study lies in the simultaneous comparison of the capabilities of several machine learning algorithms, including Random Forest (RF), Gradient Tree Boosting (GTB), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN), for identifying spatial patterns of flood susceptibility and determining areas prone to flood risk. In this regard, the use of the Google Earth Engine cloud-based processing environment, combined with Sentinel-1 radar data for identifying flood locations and a diverse set of topographic, hydrological, climatic, vegetation, soil, geological, and land-use factors, enables the simultaneous investigation of the factors influencing flood susceptibility. Another innovative aspect of the study is the comparative evaluation of model performance using different evaluation metrics and the identification of the model with more suitable performance for detecting high-risk areas. In addition to providing an accurate and comparative representation of the capabilities of different algorithms, this approach can provide a practical basis for identifying priority areas, planning preventive measures, and managing and reducing flood risk in the Doostbiglo watershed and similar watersheds.
Material and Methods
This research aimed to zone flood susceptibility in the Doostbiglo watershed, employing advanced machine learning algorithms within the GEE cloud platform. Initially, flood-influencing environmental factors, including lithology, topography (elevation, slope), hydrological indices (TWI, SPI), distances to rivers and roads, runoff, rainfall, vegetation cover (NDVI), soil texture, and land use, were collected and processed using reliable sources and the GEE environment. Subsequently, all layers were fuzzyfied to determine the relative importance of each factor in the modeling process. For model training, 585 flood points and 699 non-flood points were collected using Sentinel-1 satellite imagery. The data were split into a 70% training set and a 30% testing set. Four machine learning algorithms, including Random Forest (RF), Gradient Boosting Trees (GTB), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN), were implemented and trained in GEE. Finally, the models’ performance was evaluated using statistical metrics ROC and AUC to identify the best model for the final flood susceptibility zoning of the targeted watershed.
Results and Discussion
This research, utilizing Sentinel-1 radar imagery, successfully identified and extracted 585 flood points and 699 non-flood points in the study area, which served as the basis for subsequent analyses. The analysis of factors influencing flood occurrence in the Doostbiglo watershed revealed complex spatial patterns. The northern and eastern parts of the watershed, due to their low elevation and slope, along with high values of Topographic Wetness Index (TWI), were identified as susceptible to runoff accumulation. Conversely, the western, eastern border, and southeastern regions of the watershed are exposed to severe erosion and runoff due to high rainfall intensity and steep slopes; this condition is exacerbated by the Stream Power Index (SPI), which confirms high erosive power. The high density of streams in the north, east, and southwest of the region plays a key role in guiding and concentrating water flow and increasing flood risk. Furthermore, factors such as road density (as impervious surfaces) in the central and northwestern areas, sparse vegetation cover in the west and north (reducing soil absorption), low soil permeability (including clay loam textures and marl, shale, and conglomerate rocks), and specific land uses (aquatic, urban, rural, and agricultural areas) have led to increased surface runoff and intensified the destructive effects of floods in these areas.
Conclusion
During the modeling phase using machine learning algorithms, the importance of various factors in flood prediction was examined. Among these, slope was identified as the most influential factor in the machine learning process, accounting for 12.24% of the importance. Land use (11.98%) and vegetation cover (11.88%) ranked next in importance. The RF, GTB, SVM, and KNN algorithms generally classified the eastern and northern areas, as well as parts of the western slopes of Sabalan, into high and very high flood risk categories. Based on the models' estimates, the extent of high-risk flood areas constituted 21.42%, 21.94%, 21.31%, and 18.08% of the total study area, respectively. In the performance evaluation, the Gradient Tree Boosting (GTB) algorithm achieved the highest performance in distinguishing between flood-prone and non-flood-prone areas, with an AUC value of 0.965. The Random Forest (RF) algorithm followed closely with an AUC of 0.964. Additionally, the SVM and KNN algorithms achieved AUC values of 0.926 and 0.920, respectively. Overall, all machine learning algorithms demonstrated satisfactory performance in zoning flood susceptibility within the Doost-Bigloo watershed; however, the GTB model outperformed the other models in distinguishing between flood-prone and non-flood-prone areas due to its highest AUC value.
Keywords
Subjects
Rahmati, M., Balenzano, A., Bechtold, M., Brocca, L., Fluhrer, A., Jagdhuber, T., … & Montzka, C. (2026). Soil moisture retrieval from Sentinel-1: Lessons learned after more than a decade in orbit. Remote Sensing of Environment, 333, 115146. https://doi.org/10.1016/j.rse.2025.115146
Siddiquie, K. T., & Nafi, A. M. (2025). Flood hazard zonation using 2D HEC-RAS and explainable machine learning in urban watersheds. Research Journal in Civil, Industrial and Mechanical Engineering, 2(2), 112–130. https://doi.org/10.61424/rjcime.v2i2.631
