Detection of temporarily flooded vegetation using time series of dual polarised C-band synthetic aperture radar data

Kurzfassung

The intense research of the last decades in the field of flood monitoring has shown that microwave sensors provide valuable information about the spatial and temporal flood extent. The new generation of satellites, such as the Sentinel-1 (S-1) constellation, provide a unique, temporally high-resolution detection of the earth's surface and its environmental changes. This opens up new possibilities for accurate and rapid flood monitoring that can support operational applications. Due to the observation of the earth's surface from space, large-scale flood events and their spatiotemporal changes can be monitored. This requires the adaptation of existing or the development of new algorithms, which on the one hand enable precise and computationally efficient flood detection and on the other hand can process a large amounts of data. In order to capture the entire extent of the flood area, it is essential to detect temporary flooded vegetation (TFV) areas in addition to the open water areas. The disregard of temporary flooded vegetation areas can lead to severe underestimation of the extent and volume of the flood. Under certain system and environmental conditions, Synthetic Aperture Radar (SAR) can be utilized to extract information from under the vegetation cover. Due to multiple backscattering of the SAR signal between the water surface and the vegetation, the flooded vegetation areas are mostly characterized by increased backscatter values. Using this information in combination with a continuous monitoring of the earth's surface by the S-1 satellites, characteristic time series-based patterns for temporary flooded vegetation can be identified. This combination of information provides the foundation for the time series approach presented here. This work provides a comprehensive overview of the relevant sensor and environmental parameters and their impact on the SAR signal regarding temporary open water (TOW) and TFV areas. In addition, existing methods for the derivation of flooded vegetation are reviewed and their benefits, limitations, methodological trends and potential research needs for this area are identified and assessed. The focus of the work lies in the development of a SAR and time series-based approach for the improved extraction of flooded areas by the supplementation of TFV and on the provision of a precise and rapid method for the detection of the entire flood extent. The approach developed in this thesis allows for the precise extraction of large-scale flood areas using dual-polarized C-band time series data and additional information such as topography and urban areas. The time series features include the characteristic variations (decrease and/or increase of backscatter values) on the flood date for the flood-related classes compared to the whole time series. These features are generated individually for each available polarization (VV, VH) and their ratios (VV/VH, VV-VH, VV+VV). The generation of the time series features was performed by Z-transform for each image element, taking into account the backscatter values on the flood date and the mean value and standard deviation of the backscatter values from the nonflood dates. This allowed the comparison of backscatter intensity changes between the image elements. The time series features constitute the foundation for the hierarchical threshold method for deriving flood-related classes. Using the Random Forest algorithm, the importance of the time series data for the individual flood-related classes was analyzed and evaluated. The results showed that the dual-polarized time series features are particularly relevant for the derivation of TFV. However, this may differ depending on the vegetation type and other environmental conditions. The analyses based on S-1 data in Namibia, Greece/Turkey and China during large-scale floods show the effectiveness of the method presented here in terms of classification accuracy. Theiv supplementary integration of temporary flooded vegetation areas and the use of additional information resulted in a significant improvement in the detection of the entire flood extent. It could be shown that a comparably high classification accuracy (~ 80%) was achieved for the flood extent in each of study areas. The transferability of the approach due to the application of a single time series feature regarding the derivation of open water areas could be confirmed for all study areas. Considering the seasonal component by using time series data, the seasonal variability of the backscatter signal for vegetation can be detected. This allows for an improved differentiation between flooded and non-flooded vegetation areas. Simultaneously, changes in the backscatter signal can be assigned to changes in the environmental conditions, since on the one hand a time series of the same image element is considered and on the other hand the sensor parameters do not change due to the same acquisition geometry. Overall, the proposed time series approach allows for a considerable improvement in the derivation of the entire flood extent by supplementing the TOW areas with the TFV areas.