Abstract's details
Deep Learning and SAR Altimetry Techniques in Coastal Island Areas
Event: 2022 Ocean Surface Topography Science Team Meeting
Session: Coastal Altimetry
Presentation type: Type Poster
Contribution: not provided
Abstract:
Synthetic Aperture Radar (SAR) Altimetry has made a remarkable progress over the past years. Advances in data processing, combined with technological progress such as the advent of new Altimetry satellites (Sentinel 3A,3B,6, SWOT etc.) increased the accuracy of the retrieved geophysical parameters (i.e., Sea Level Anomaly, Significant Wave Height and Wind Speed) in coastal zones within several hundred meters from the coastline.
The improvement in the estimation of the geophysical parameters using SAR Altimetry has been reported by many researchers. The improved accuracy is obtained through the development of new SAR Altimetry retracking algorithms in several research and development projects (i.e., SAR Altimetry Mode Studies and Applications- SAMOSA). Similar to Low Resolution Mode (LRM) Altimetry, the requirement of specialised retrackers for SAR waveforms is vital in improving the estimated ocean parameters. The waveform retracking is a postprocessing protocol to convert waveforms into scientific parameters of power amplitude (related to wind speed), range (related to sea level), and slope of leading edge (related to SWH) that characterise the observed scene (Idris et al., 2021).
However, several issues remain open. Close to the coastline, SAR altimeter simultaneously views scattering surfaces of both water and land producing complicated waveform patterns therefore a huge range of waveform shapes is observed. This complexity poses a real challenge to today’s approach to retrack waveform.
This work aims to present results from an in-house developed Deep Learning algorithm, in order to retrack waveforms by learning complicated patterns in coastal areas. Initially various steps of data preparations have to be conducted to receive the waveforms leading edge position:
1. Normalisation of the waveforms
2. Labelling of the waveforms
3. The training, developing and testing data sets
The use of a Convolutional Neural Networks will be introduced. With CNNs the use of waveforms images will be feasible, therefore the algorithm will be able to process more waveforms of a satellite track and provide more information instead of taking into account one waveform or a single part of it.
Because we are working with time series, it is beneficial to know the previous assumed leading edge position for the current case. By using Recurrent Neural Networks (RNN), it is possible to give the information of the assumed actual leading edge position to the next following waveform analysis. However, of course, it is also feasible to combine the CNN and RNN (Mattes, 2019).
REFERENCES
Idris N., Vignudelli S., Deng X., (2021) Assessment of retracked sea levels from Sentinel-3A Synthetic Aperture Radar (SAR) mode altimetry over the marginal seas at Southeast Asia, International Journal of Remote Sensing, 42:4, 1535-1555, DOI: 10.1080/01431161.2020.1836427
Mattes D.,(2019) Analysis of Waveforms in the Satellite Altimetry by Using Neural Networks, Stuttgart University
The improvement in the estimation of the geophysical parameters using SAR Altimetry has been reported by many researchers. The improved accuracy is obtained through the development of new SAR Altimetry retracking algorithms in several research and development projects (i.e., SAR Altimetry Mode Studies and Applications- SAMOSA). Similar to Low Resolution Mode (LRM) Altimetry, the requirement of specialised retrackers for SAR waveforms is vital in improving the estimated ocean parameters. The waveform retracking is a postprocessing protocol to convert waveforms into scientific parameters of power amplitude (related to wind speed), range (related to sea level), and slope of leading edge (related to SWH) that characterise the observed scene (Idris et al., 2021).
However, several issues remain open. Close to the coastline, SAR altimeter simultaneously views scattering surfaces of both water and land producing complicated waveform patterns therefore a huge range of waveform shapes is observed. This complexity poses a real challenge to today’s approach to retrack waveform.
This work aims to present results from an in-house developed Deep Learning algorithm, in order to retrack waveforms by learning complicated patterns in coastal areas. Initially various steps of data preparations have to be conducted to receive the waveforms leading edge position:
1. Normalisation of the waveforms
2. Labelling of the waveforms
3. The training, developing and testing data sets
The use of a Convolutional Neural Networks will be introduced. With CNNs the use of waveforms images will be feasible, therefore the algorithm will be able to process more waveforms of a satellite track and provide more information instead of taking into account one waveform or a single part of it.
Because we are working with time series, it is beneficial to know the previous assumed leading edge position for the current case. By using Recurrent Neural Networks (RNN), it is possible to give the information of the assumed actual leading edge position to the next following waveform analysis. However, of course, it is also feasible to combine the CNN and RNN (Mattes, 2019).
REFERENCES
Idris N., Vignudelli S., Deng X., (2021) Assessment of retracked sea levels from Sentinel-3A Synthetic Aperture Radar (SAR) mode altimetry over the marginal seas at Southeast Asia, International Journal of Remote Sensing, 42:4, 1535-1555, DOI: 10.1080/01431161.2020.1836427
Mattes D.,(2019) Analysis of Waveforms in the Satellite Altimetry by Using Neural Networks, Stuttgart University