Abstract's details
A near global improved gridded multi-mission daily SLA product slightly beyond real time..
CoAuthors
Event: 2023 Ocean Surface Topography Science Team Meeting
Session: Application development for Operations (ROUND TABLE)
Presentation type: Type Poster
Contribution: PDF file
Abstract:
In order to cover data gaps and mitigate data errors in gridded satellite altimetry both MEaSURES and CMEMSa global products are created using large temporal windows during interpolation, as well as spatial and temporal filtering. In this presentation we present a new product from the Technical University of Denmark (DTU) being a new global ¼ ⁰ x ¼ ⁰ SLA grid product containing more well-resolved features in time. Compared to related SLA data products from CMEMS and MEaSUREs, the SLA grid created in this project performs similarly. The DTU products is updated daily with less than 24 hours latency.
As an alternative to traditional filtering methods, an autoencoder is applied to the data after gridding to create a smart filter for removing satellite tracks and non-physical noise. Preliminary results are propitious, as satellite tracks are successfully removed from the data without influencing the major underlying features. The size of the latent space in the autoencoder directly determines the amount of information propagated through the filter. Finally, the data product is used to forecast near-future values of the SLA grid using an Encoder-Decoder Attention Convolutional Long Short-Term Memory (EDAConvLSTM) architecture to close the publishing gap created by the moving temporal window in the gridding process. The prediction power of the model is very dependent on the smoothness of the data. For the SLA grid product produced in this report, however, the forecast may be used to extend the grid product slightly beyond real-time, such that a few future days are included.
As an alternative to traditional filtering methods, an autoencoder is applied to the data after gridding to create a smart filter for removing satellite tracks and non-physical noise. Preliminary results are propitious, as satellite tracks are successfully removed from the data without influencing the major underlying features. The size of the latent space in the autoencoder directly determines the amount of information propagated through the filter. Finally, the data product is used to forecast near-future values of the SLA grid using an Encoder-Decoder Attention Convolutional Long Short-Term Memory (EDAConvLSTM) architecture to close the publishing gap created by the moving temporal window in the gridding process. The prediction power of the model is very dependent on the smoothness of the data. For the SLA grid product produced in this report, however, the forecast may be used to extend the grid product slightly beyond real-time, such that a few future days are included.