Mr. Aurélien Colin (March 3, 2021, 17:30-18:00)
|Affiliation||IMT Atlantique & CLS|
|Title||Semantic Segmentation of Metocean Processes and Estimation of Ancillary Data|
Operating despite cloud cover and unaffected by night, SAR imagery is particularly useful to study metocean phenomena. However, data gathering is costly as it requires expert knowledge to accurately design the various phenomena occurring on the ocean surface. The TenGeoP-SARwv dataset provides categorical information for ten metocean processes and can be used in a weakly supervised framework to provide pseudo-segmentation of these phenomena. Manual segmentation of the SAR observations prove to be even more successful even with a very low amount of ground truths. This presentation shows that the segmentation of the metocean processes is achievable on Wave Mode images and can be used on wider Wide Swath observations. The resulting segmentations can be enhanced by the estimation of ancillary variables, such as the reflectivity of weather radars, to further increase the knowledge of the meteocean situation.
Dr. Hirotaka Hachiya (March 3, 2021, 18:00-18:30)
|Title||Spatio-temporal integration of forecast guidance outputs using U-Net|
Recently, serious disasters have been caused by heavy rainfall in Japan. Currently, three 2-D precipitation forecasts, called LFM, MSM, and GSM are separately operated and integrating them is expected to improve the accuracy of the forecast. So far, arithmetic and Bayesian average methods have been actively studied, but in these methods, spatial dependency for the integration is not taken into account. In addition, extra information could not be utilized in a straightforward manner. To migrate these limitations, we propose to define the problem of integrating forecasts as the problem of image-to-image transformation, where forecasts images are converted to weight ratio images. The proposed method extending U-Net enables us to generate pixel (location)-wise weights and seamlessly to embed extra information and forecast times. The effectiveness of our proposed method is shown through experiments with precipitation forecast tasks.