مهيار گرشاسبي

Floods represent one of the most devastating natural calamities globally, with their impact escalating due to factors like population expansion and climate fluctuations. Precise flood mapping stands as a crucial measure in mitigating the consequences of this phenomenon. Leveraging remote sensing data, particularly synthetic aperture radar (SAR), has revolutionized flood mapping, enabling the assessment of extensive flood-affected areas. The integration of SAR data with data-driven computational tools, such as artificial intelligence (AI) techniques, has significantly enhanced the accuracy of satellite image processing. Nonetheless, flood delineation through SAR encounters inherent challenges, and inaccuracies in classification can yield irreversible repercussions. Hence, incorporating quantitative uncertainties as interpretable probabilities becomes imperative to facilitate risk-informed decision-making. This study presents comparative findings between the Standard U-net model and the proposed Bayesian U-net model, the latter being a probabilistic deep learning approach, for flood mapping utilizing SAR data from the freely available Sentinel-1 sensor. Through rigorous eva‎luation employing a diverse array of performance metrics, we scrutinize the disparities between the two models. Notably, the results indicate superior performance of the Bayesian U-net model across various metrics including Accuracy, Intersection over unio‎n (IoU), F1-score, and Recall. These outcomes underscore not only the heightened accuracy achieved by the proposed model but also its efficacy in quantifying uncertainty associated with flood mapping, thereby advocating for its adoption in practical flood management scenarios.

نقشه‌برداري سيل , روش‌هاي بيزي , يادگيري‌عميق

Flood Mapping , Bayesian methods , Deep Learning

Mahyar Garshasbi

Dr. Hossein Alizadeh, Dr. Motahare Saadatpour