hide
Free keywords:
-
Abstract:
Flood is one of the most widespread and frequent natural disasters. Deriving accurate and rapid cartographic information on flood extent is essential to help manage the situation. Satellite remote sensing is now
widely used for near real-time flood monitoring as it provides large scale detection in a time- and cost-efficient manner. Optical satellite imagery is employed as important tools for flood mapping due to easier
interpretability and high spatial resolution. However, cloudy weather associated with floods are a great
obstacle to optical sensors for flood monitoring. In contrast, Synthetic Aperture Radar (SAR) allows observation of wide areas across near all-weather conditions and plays a significant role in operational services
for flood management. Although in many cases smooth water surfaces can be easily extracted from SAR
imagery, it is subjected to overestimation of flooded areas especially in the arid and semi-arid regions since
the complex interactions between SAR characteristics and environmental conditions.
Advanced machine learning and deep learning approaches have demonstrated large potential to overcome
the problem by learning features directly from images which requires a large number of labeled samples for
training and validation. Therefore, some public georeferenced dataset to train and test deep learning flood
algorithms are being produced.
To investigate the role of globally available label datasets in obtaining reliable flood maps using SAR data
and deep learning approaches, we tried one of the open access dataset, Sen1Floods11, which is a surface
water dataset. We trained, validated and tested a ResNet50 model to segment flood water using a subset
of this dataset.The classification results of flood water have obtained an overall accuracy of 89.5% for the
test dataset in India and 78.9% for the test dataset in Pakistan. Results show the potential of the flood water
dataset to better detect the flooded area.
