Intense rainfall estimation using radar reflectivity feature extraction for flood forecasting

This thesis investigates the use of weather radar data for hydrological applications, specifically in rainfall estimation for flood forecasting, focusing on two major river basins in Selangor, Malaysia. Weather radar provides high-resolution rainfall estimates over large areas, addressing problems such as uneven rainfall distribution and limited rain gauge coverage. However, radar-based Quantitative Precipitation Estimation (QPE) is susceptible to errors caused by ground clutter, beam blockage, and signal attenuation. The research explores innovative techniques in radar QPE accuracy enhancement, including using artificial intelligence tools to classify rain types and improve the estimated rainfall values. Methodologies include the extraction of geometric characteristics from radar reflectivity images, including rain cell location, area, intensity, and orientation, which are then used for rainfall classification and machine learning applications. The K-Nearest Neighbor (KNN) algorithm is applied to classify radar-based rainfall into convective and stratiform categories. Despite some misclassifications due to overlapping feature characteristics, the KNN model achieves strong predictive performance with an Area Under the Curve (AUC) greater than 90%. In addition, an Artificial Neural Network (ANN) model is designed using the Levenberg-Marquardt algorithm and rain gauge data to improve the accuracy of radar-based QPE. The performance of the ANN-enhanced QPE model is evaluated using Root Mean Square Error (RMSE) and correlation metrics, showing a significant reduction in RMSE from 11.40 mm/h to 4.76 mm/h, thus enhancing the precision of radar-based rainfall estimates. The improved radar QPE model is then applied as input to an integrated flood forecasting system, employing the HEC-HMS hydrological model, with a novelty of merging process of the gridded-based QPE as sub-basin rainfall. The integrated system is applied for flood simulation in the Klang and Langat River Basins, with calibration results demonstrating that rain gauge data and ANN-calibrated QPE produces the highest correlation followed by uncalibrated radar QPE. The integration of calibrated radar QPE with hydrological models presents a promising approach for improving flood forecasting and predicting extreme weather events such as flash floods. This research contributes to advancements in radar hydrology, hybrid machine learning techniques for feature extraction, enhanced rainfall estimation methods, and improved radar-based rainfall-runoff models for better flood forecasting and warning systems in river basins in Malaysia.