9:15 AM - 9:30 AM
[MIS10-02] Development of AMSR2 high-resolution thin ice thickness estimation with advanced techniques for coastal polynya detection
Keywords:Thin ice thickness, Coastal polynya, AMSR2
Coastal polynya (thin ice areas and low ice concentration areas) are areas of high ice production due to huge heat loss to the atmosphere. The release of cold and saline brine due to the ice production leads to the formation of Dense Shelf Water (DSW). The DSW serves as the source water for Antarctic Bottom Water in the Southern Ocean and North Pacific Intermediate Water in the Sea of Okhotsk. Moreover, in the Arctic Ocean and the Bering Sea, DSW plays a crucial role in the formation and maintenance of the Cold Halocline Layer. Therefore, estimating sea ice production in coastal polynyas is essential as foundational information for understanding processes such as deep/intermediate water formation, ocean stratification, and biogeochemical cycles such as CO2.
Many studies have developed algorithms to derive thin ice thickness from satellite microwave radiometer data and have subsequently estimated sea ice production based on heat flux calculation using the thin ice thickness. These algorithms detect thin ice thickness empirically from the polarization ratio associated with sea ice thickness obtained from thermal infrared sensor such as MODIS. However, there are still areas for improvement: (1) The algorithm uses raw brightness temperature (TB) data with stripe noises and blur directly as input for the empirical equation. (2) These algorithms do not sufficiently account for the influence of the atmosphere on TB. (3) A simple regression analysis for a single variable, using only thin ice pixels, is employed. This study aims to create a high-quality thin ice thickness product by addressing these issues through noise removal and resolution enhancement of TB data, atmospheric effects reduction, and utilizing a high expressive estimation model with multiple variables. This is expected not only to improve the accuracy of sea ice production estimates in coastal polynyas, but also enable estimates in narrow polynyas and leads that were previously undetectable.
In this study, we used TBs at 19GHz, 36GHz and 89GHz from GCOM-C/AMSR2 Level-1B product. We firstly used a de-striping technique proposed by Bouali and Ladjal (2011) to remove stripe noise from the TBs at 89GHz. Subsequently, we utilized a resolution enhancement technique proposed by Long et al. (1993) to downscale the TBs at 18GHz, 36GHz, and 89GHz to grid sizes of 10km, 5km, and 2.5km, respectively. Next, we used ERA5 re-analysis dataset to remove the effect of water vapor based on a simple radiative transfer model. Finally, we trained a neural network model with three hidden layers using these correct TBs and MODIS images with a grid spacing of 2.5 km as input and ground truth, respectively. To control false detections of thin ice, we employed a multi-output architecture taking account of two types of losses: mean squared error for thin ice thickness estimation and binary cross-entropy for thin ice detection.
For a previous algorithm that estimates thin ice thickness at a resolution of 12.5 km, a false detection rate of thin ice and thick ice are calculated to be 41% and 27%, and root mean squared error (RMSE) is calculated to be 7.4 cm. Our thin ice thickness estimation model reduces to the false detection rate of 33% and 25%, and an RMSE of 5.5 cm. The result means that our proposed model achieves both higher spatial resolution and higher accuracy in thin ice thickness estimates. Also, comparison with MODIS images demonstrates improvements in various aspects, including reduced false-detection of thin ice due to atmospheric effects, decreased false detection in landfast ice, and enhanced detection for narrow polynyas and leads.
Reference:
Bouali, M. and Ladjal, S., Toward optimal destriping of MODIS data using a unidirectional variational model, IEEE Trans. Geosci. Remote Sens, 49(8), 2924-2935, 2011.
Long, D. G., Hardin, P. J., and Whiting, P. T., Resolution enhancement of spaceborne scatterometer data, IEEE Trans. Geosci. Remote Sens., 31, 700-715, 1993.
Many studies have developed algorithms to derive thin ice thickness from satellite microwave radiometer data and have subsequently estimated sea ice production based on heat flux calculation using the thin ice thickness. These algorithms detect thin ice thickness empirically from the polarization ratio associated with sea ice thickness obtained from thermal infrared sensor such as MODIS. However, there are still areas for improvement: (1) The algorithm uses raw brightness temperature (TB) data with stripe noises and blur directly as input for the empirical equation. (2) These algorithms do not sufficiently account for the influence of the atmosphere on TB. (3) A simple regression analysis for a single variable, using only thin ice pixels, is employed. This study aims to create a high-quality thin ice thickness product by addressing these issues through noise removal and resolution enhancement of TB data, atmospheric effects reduction, and utilizing a high expressive estimation model with multiple variables. This is expected not only to improve the accuracy of sea ice production estimates in coastal polynyas, but also enable estimates in narrow polynyas and leads that were previously undetectable.
In this study, we used TBs at 19GHz, 36GHz and 89GHz from GCOM-C/AMSR2 Level-1B product. We firstly used a de-striping technique proposed by Bouali and Ladjal (2011) to remove stripe noise from the TBs at 89GHz. Subsequently, we utilized a resolution enhancement technique proposed by Long et al. (1993) to downscale the TBs at 18GHz, 36GHz, and 89GHz to grid sizes of 10km, 5km, and 2.5km, respectively. Next, we used ERA5 re-analysis dataset to remove the effect of water vapor based on a simple radiative transfer model. Finally, we trained a neural network model with three hidden layers using these correct TBs and MODIS images with a grid spacing of 2.5 km as input and ground truth, respectively. To control false detections of thin ice, we employed a multi-output architecture taking account of two types of losses: mean squared error for thin ice thickness estimation and binary cross-entropy for thin ice detection.
For a previous algorithm that estimates thin ice thickness at a resolution of 12.5 km, a false detection rate of thin ice and thick ice are calculated to be 41% and 27%, and root mean squared error (RMSE) is calculated to be 7.4 cm. Our thin ice thickness estimation model reduces to the false detection rate of 33% and 25%, and an RMSE of 5.5 cm. The result means that our proposed model achieves both higher spatial resolution and higher accuracy in thin ice thickness estimates. Also, comparison with MODIS images demonstrates improvements in various aspects, including reduced false-detection of thin ice due to atmospheric effects, decreased false detection in landfast ice, and enhanced detection for narrow polynyas and leads.
Reference:
Bouali, M. and Ladjal, S., Toward optimal destriping of MODIS data using a unidirectional variational model, IEEE Trans. Geosci. Remote Sens, 49(8), 2924-2935, 2011.
Long, D. G., Hardin, P. J., and Whiting, P. T., Resolution enhancement of spaceborne scatterometer data, IEEE Trans. Geosci. Remote Sens., 31, 700-715, 1993.