Mapping flooded areas using Sentinel-1 radar satellite imagery series through Machine learning and Deep learning methods

                Rapid and accurate mapping of floods offers an excellent advantage for local, regional decision-makers to mitigate the exposure of human and material losses. The current study assessed the performance of several machine learning (ML) and deep learning (DL) models for detecting and mapping floods using Sentinel-1 SAR imagery. Three distinct approaches were used with machine learning and deep learning models: pixel classification, object-based image analysis and object instance segmentation. The ML models are Random Forest (RF) and Support Vector Machine applied for pixel classification and object-based image analysis. The DL models are U-NET, DeepLabV3 and Mask RCNN used for pixel classification and object instance segmentation. The models were implemented using SNAP (Sentinel Application Platform), ESRI ArcGIS Pro, Esri ArcGIS API for Python and Python programming language. To test the model's scalability, five different cases studies were selected. These areas are located in Romania (Prut River sector, Timiș River sector and Râul Negru sector), the United States of America (Missouri River sector) and Australia (Broughton Creek sector). Five Sentinel-1 images were used for each flood, having four collected previous to the flood event and one collected after the flood event. Each Sentinel-1 image was calibrated and ortho-corrected, and filtered using SNAP. The intensity images were stacked and scaled in the range of the intensity thresholds associated with water and non-water so that all the case studies have the same margins for intensity. Further, samples were collected in ArcGIS Pro from the Prut River region using the stack of images created from the previous step. Besides water, other classes, such as forest, agricultural fields and bare soil, were collected and used in the training process. The training for the ML models took place directly on the standardized radar images within ArcGIS Pro. The training of the DL models was done through the use of Jupyter Notebooks and ArcGIS API for Python. The models were trained on samples collected from the Prut River area and then tested on all selected regions to assess their ability to perform in different study areas. The highest accuracy, calculated as Intersect over Union, was obtained by the U-Net model (IoU score of 0.74). Comparable accuracies were obtained by the RF and SVM models implemented with OBIA, with an IoU score of 0.72. Mask R-CNN and DeepLabV3 got IoU scores of 0.70, and the lowest accuracies for floods mapping were obtained by the RF and SVM models implemented as pixel classification (both having IoU scores of 0.53).