The aim of this study was to find the most capable classification / detection model, which can be used in combination with a series of Sentinel-1GRD radar satellite images to map floods as quickly and accurately as possible. The most suitable frameworks for this task were SNAP and ArcGIS Pro, which offer numerous and powerful geoprocessing tools, thus being able to achieve a workflow in which I integrated all the necessary steps of my initial goal. Therefore, I chose a series of the most well-known and used machine learning and deep learning algorithms that I could find and I was not disappointed by the combination of machine learning algorithms with object-oriented analysis, nor by the most promising algorithm of the moment, namely U-Net. The first necessary step was to standardize the radar images for both the training region and ones used to test the accuracy of the models. The next step in the workflow was sampling, which differs from deep learning models, in that the machine learning models, also require sampling of other environmental components beside water, such as forests, agricultural fields and bare soil, in order to be able to correctly classify the rasters. Then, the training for the established machine learning models took place directly on the standardized radar images within ArcGIS Pro. Meanwhile, the training of convolutional neural networks took place through an ArcGIS library in Python, on the image chips. After the training of the models was done, I proceeded to the actual testing of the models and the results varied depending on the region, the size of the floods and the technique that I used. As a consequence of the results I obtained in this study, I consider that the immediate mapping of extreme meteorological events, such as floods, can be done without special human involvement, through the use of modern techniques, which are capable of a rapid and trustworthy assessment of the situation on the ground.
Mots clés : Remote sensing|Floods|Machine learning|Deep learning|SAR
A102649AT