The determination of the wind speed from the radar backscatter values of the SAR images using the available Geophysical Model Functions requires the knowledge of the wind direction, besides the radar geometry. It is thus important to have a method to determine the wind direction directly for the SAR image, taking advantage of its high resolution to obtain more detailed wind fields, reflecting the natural variability of the wind direction. A novel method, developed in Zanchetta and Zecchetto, 2021, to obtain high resolution (up to 500 m) wind fields from SAR images is here illustrated. It employs a convolutional neural network (Goodfellow, 2016) with a residual network structure (ResNet) (He, 2016), specifically designed to obtain the aliased wind direction without any other external information, used only to solve the 180° ambiguity. This method has been applied to Sentinel-1 SAR images in coastal areas: forty-seven SAR images have been processed with the ResNet, previously trained with other twenty-file images. The obtained wind fields reproduce successfully the meteorological situations characterized by strong divergence, independently of atmospheric stability conditions and, remarkably, even in the absence of wind streaks. Furthermore, they do not seem influenced by the presence of convective turbulence structures, atmospheric lee waves, ships and platforms. (Fig.1) Statistical analysis was carried out comparing the SAR-derived wind directions with those from ECMWF atmospheric model, ASCAT scatterometer and in-situ gauges. It reports biases ? equal to -1.1o, 2.4o and -4.6o respectively, and centered root mean square difference cRMSd < 21o, consistent with the benchmark obtained comparing scatterometer with ECMWF wind directions over the areas imaged by SAR (? = 2.1o, cRMSd = 19o). Examples of SAR-derived wind fields are illustrated. Some particular situations clearly showing the new opportunities of investigating the wind field offered by this method are also presented, including applications in enclosed basins, in close proximity of the coastline, and in presence of complex flow situations. References Goodfellow, I., Bengio, Y. and Courville, A., Deep Learning, MIT Press, 2016. He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778. Zanchetta, A. and S. Zecchetto, Wind direction retrieval from Sentinel-1 SAR images using ResNet, Remote Sensing of Environment, 253, February 2021, doi: 10.1016/j.rse.2020.112178

Wind direction retrieval from SAR images using ResNet

2021

Abstract

The determination of the wind speed from the radar backscatter values of the SAR images using the available Geophysical Model Functions requires the knowledge of the wind direction, besides the radar geometry. It is thus important to have a method to determine the wind direction directly for the SAR image, taking advantage of its high resolution to obtain more detailed wind fields, reflecting the natural variability of the wind direction. A novel method, developed in Zanchetta and Zecchetto, 2021, to obtain high resolution (up to 500 m) wind fields from SAR images is here illustrated. It employs a convolutional neural network (Goodfellow, 2016) with a residual network structure (ResNet) (He, 2016), specifically designed to obtain the aliased wind direction without any other external information, used only to solve the 180° ambiguity. This method has been applied to Sentinel-1 SAR images in coastal areas: forty-seven SAR images have been processed with the ResNet, previously trained with other twenty-file images. The obtained wind fields reproduce successfully the meteorological situations characterized by strong divergence, independently of atmospheric stability conditions and, remarkably, even in the absence of wind streaks. Furthermore, they do not seem influenced by the presence of convective turbulence structures, atmospheric lee waves, ships and platforms. (Fig.1) Statistical analysis was carried out comparing the SAR-derived wind directions with those from ECMWF atmospheric model, ASCAT scatterometer and in-situ gauges. It reports biases ? equal to -1.1o, 2.4o and -4.6o respectively, and centered root mean square difference cRMSd < 21o, consistent with the benchmark obtained comparing scatterometer with ECMWF wind directions over the areas imaged by SAR (? = 2.1o, cRMSd = 19o). Examples of SAR-derived wind fields are illustrated. Some particular situations clearly showing the new opportunities of investigating the wind field offered by this method are also presented, including applications in enclosed basins, in close proximity of the coastline, and in presence of complex flow situations. References Goodfellow, I., Bengio, Y. and Courville, A., Deep Learning, MIT Press, 2016. He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778. Zanchetta, A. and S. Zecchetto, Wind direction retrieval from Sentinel-1 SAR images using ResNet, Remote Sensing of Environment, 253, February 2021, doi: 10.1016/j.rse.2020.112178
2021
Deep residual network
Synthetic Aperture Radar (SAR)
Wind field
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14243/418782
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact