High spatial and temporal resolution flood monitoring through integration of multisensor remotely sensed data and Google Earth Engine processing

Climate change and anthropogenic impact are intensifying the frequency and intensity of extreme flood events. This is particularly worrying in the Mediterranean area, which is highly vulnerable and therefore subject to increased flood risk. The monitoring of flooded areas at high-resolution plays an important role in all phases of disaster management, from alert to the emergency and civil protection phase, up to damage assessment, for compensation and risk reduction purposes.

This study aims at the multi-temporal analysis of remote sensing data, mainly radar data, through the implementation of a semi-automated system for the high-resolution mapping of river flooding effects. The objective is also to develop a system based on the fusion of different data sources and for different land cover types. The system includes an algorithm for the computation of multi-temporal, probabilistic flood maps, based on the analysis of amplitude series (in dB) of a stack of SAR images, acquired both in areas with permanent water and in areas with potential flooding. Exploiting a Bayesian inference framework, conditioned probabilities are estimated for the presence of water. The procedure relies on the temporal modelling of the SAR amplitudes time series, in order to account for seasonal and other slow temporal trends, and thus highlighting floods as events causing abrupt variations of the backscatter, lasting for a single or a few acquisitions. The methodology is particularly suited to data from sensors characterized by a high temporal frequency, such as the European Sentinel-1 constellation, whose two sensors acquire with the same geometrical configuration every 6 days over Europe. In parallel, a land use classification, at high resolution, is produced for each year within the period of acquisition of the satellite image stack (late 2014 to present) using Google Earth Engine [1]. This cloud-based platform makes it easy to access high-performance computing resources for processing geospatial data, allowing for the independent development of algorithms and subsequently specific applications. This supervised classification, achieved with the 'random forest' machine learning technique, is obtained through the combined use of SAR Sentinel 1 and optical Sentinel 2 images, over each entire year of interest. We show how the combination of these techniques can help gaining insight on the land cover, and on the expected changes of their appearance in the remotely sensed data in flooded conditions. This information can be used to improve the performance of the monitoring algorithm over various land cover scenarios and climatic settings.

The procedure is tested over the Metaponto plain, in the Basilicata region (southern Italy). The proposed methodologies can however be used for other contexts affected by similar events, in the Mediterranean area and worldwide.

References

• Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R. (2017) - Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, Volume 202, 2017, Pages 18-27, ISSN 0034-4257, https://doi.org/10.1016/j.rse.2017.06.031.