Abstract
The rising frequency and intensity of floods pose risks to human lives, infrastructure, and ecosystems, particularly in coastal regions, as traditional flood management systems struggle with uncertainties, complex environmental factors, and rapid urbanization, reducing decision-making accuracy. The study employs remote sensing data and a Convolutional Neural Network (CNN) to assess flood probability and risk in Port St Johns, South Africa, utilizing thirteen flood-influencing variables to minimize overfitting and extract robust features, addressing complex terrain and climate variability. The study uses data from ALOS DEM, CHIRPS, and Copernicus to analyze various factors such as Height Above the Nearest Drainage (HAND), TWI, MNDWI, TRI, distance to river, elevation, slope, aspect, curvature, flow accumulation and direction, precipitation, and land cover, using optimized kernel sizes, Rectified Linear Unit (ReLu), and regularization techniques. The results reveal significant correlations between terrain-related and hydrological factors, such as slope (3.98 %), HAND (3.07 %) and elevation (1.29 %), affecting water movement, accumulation, and drainage potential, with land cover (0.42 %) and precipitation (0.39 %) playing a secondary role. The CNN model for flood probability prediction reveals high accuracy and predictive performance, with a mean absolute error of 0.007 and a precision of 0.988 for flood-affected and unaffected areas. The InaSAFE analysis reveals that 26 % of Port St Johns’ population (870 people) and 34 % of structures (896 buildings) are directly affected by flooding, with high-risk zones affecting 420 people, 5.3 km of roads, and 479 buildings. The findings of the model enhance community safety and resilience to climate-induced flooding by improving flood risk prediction, optimizing evacuation, resource allocation, and disaster management through early warning systems and damage assessments.