پیش‌بینی مناطق مستعد سیلاب با رویکرد هوش مصنوعی و تلفیق داده‌های سنجش از دور در بستر Google Earth Engine و Google Colab

Extended Abstract

Background and Objectives

Floods are among the most destructive natural hazards, particularly in arid and semi-arid regions, causing extensive damage to urban areas, infrastructure, agricultural lands, and water resources. Climate change, increased extreme rainfall events, unplanned urban development, and land-use changes have intensified the frequency and severity of floods in recent decades. In this context, accurately identifying flood-prone areas and producing susceptibility maps play a crucial role in risk management, land-use planning, and damage mitigation. However, the lack of hydrological and ground-based data in many regions, including the Birjand Plain, limits the effectiveness of conventional hydrological methods. Consequently, leveraging remote sensing data, spectral indices, and machine learning algorithms has emerged as a novel, rapid, and cost-effective approach. The main objective of this study is to develop an integrated framework for flood susceptibility mapping in the Birjand Plain using the Normalized Difference Flood Index (NDFI), Google Earth Engine (GEE), and machine learning algorithms, and to evaluate and compare their performance.

Materials and Methods

In this study, satellite imagery was first processed in the Google Earth Engine environment. After applying preprocessing steps and cloud removal, the NDFI was computed. Using field surveys and historical aerial photographs from the 1960s, a binary map of flood and non-flood areas was produced. Based on this map, flood occurrence locations were extracted and used for model training and validation. Fifteen geo-environmental factors influencing flood occurrence, including topographic, hydrological, climatic, and land-cover parameters, were extracted and prepared as raster layers with a spatial resolution of 30 × 30 m. To prevent multicollinearity, correlations among factors were examined. Subsequently, four machine learning algorithms—Random Forest (RF), AdaBoost, Gradient Boosting (GB), and a hybrid RF–GB model—were applied to model flood susceptibility. Model performance was evaluated using Accuracy, Sensitivity, Specificity, F1-Score, Kappa coefficient, RMSE, and the Area Under the ROC Curve (AUC).

Results and Discussion

The results indicated that all machine learning models demonstrated acceptable capability in predicting flood-prone areas; however, differences in accuracy and stability were observed among them. The Random Forest model exhibited the best overall performance, with the highest discriminative ability (AUC = 0.9), overall accuracy (Accuracy = 0.83), and complete sensitivity in identifying flood-prone areas (Sensitivity = 1, Specificity = 0.6, F1 = 0.87, Kappa = 0.64), along with the lowest error (RMSE = 0.39). Gradient Boosting (AUC = 0.78, RMSE = 0.49, Accuracy = 0.75, Sensitivity = 1, Specificity = 0.4, F1 = 0.82, Kappa = 0.49) and AdaBoost (AUC = 0.82, RMSE = 0.43, Accuracy = 0.66, Sensitivity = 0.71, Specificity = 0.6, F1 = 0.71, Kappa = 0.31) also showed reasonable ability to identify flood-prone areas, but ranked lower in terms of balance between sensitivity and specificity. The hybrid RF–GB model, despite a relative improvement in discriminative ability (AUC = 0.88, RMSE = 0.41, Accuracy = 0.75, Sensitivity = 1, Specificity = 0.4, F1 = 0.82, Kappa = 0.43), did not demonstrate a significant advantage over the Random Forest model.Flood-prone maps were classified into five categories: very low, low, moderate, high, and very high. Spatial pattern analysis revealed that areas with high and very high susceptibility were predominantly located in low-elevation zones with gentle slopes, close to rivers, and with sparse vegetation cover, highlighting the dominant role of topographic and hydrological factors in the region’s flood generation processes.

Conclusion

The findings of this study indicate that integrating the NDFI index, remote sensing data, and machine learning algorithms within the Google Earth Engine framework provides an efficient, rapid, and cost-effective approach for predicting flood-prone areas in arid and data-scarce regions. Among the evaluated models, Random Forest demonstrated the most stable and accurate performance, showing a strong capability in identifying high-risk areas. Therefore, it can serve as a valuable decision-support tool for flood risk management, land-use planning, and mitigation of flood-related damages. The proposed framework is transferable to other regions with similar environmental conditions and can contribute to enhancing regional resilience to flood hazards.