Modifierad RCNN för flodöversvämningsdetektion i SAR-bilder : Fallstudier av tyska och italienska översvämningar

Riverine flooding is an increasing problem expected to affect 132 million people annually by 2030. Flood detection methods are crucial for developing effective flood mitigation and prevention policies, and are set to become even more important in the decades to come. While there are many methods for detecting floods, research on flood detection on bi-temporal SAR imagery using Pulse-Coupled Neural Networks (PCNNs) largely stalled in the previous decade. Building upon the work of R. Liu et al. (2016), Chen et al. (2011) and H. Liu et al. (2024) a modified Random-Coupled Neural Network (RCNN) is here introduced to handle simultaneous noise suppression and propagation of potentially flooded areas combined with an automatic parameter setting method. The model was tested on two datasets from Italy and Germany with F1-scores of approximately 0.38 and 0.35 respectively for the tiled datasets. Testing on the untiled datasets showed a marked improvement with F1-scores reaching approximately 0.63 and 0.60 respectively. Although the results can be characterised as subpar, questions exist about the comparability of the predictions to the ground truths. Regardless of this, the generalisability of the model is hampered by the manual tuning of kernel parameters. Further research is needed on a way of setting these parameters. Despite subpar performance against the ground truths and questions about generalisability, PCNNs offer a versatile platform from which to build upon.