Towards a new multivariate approach of Sea Surface Height editing: experimenting with data mining algorithms

Pierre Prandi (CLS, France)


Lionel Zawadzki (CLS, France); Annabelle Ollivier (CLS, France); Bruno Picard (CLS, France); Nicolas Picot (CNES, France)

Event: 2016 Ocean Surface Topography Science Team Meeting

Session: Regional and Global CAL/VAL for Assembling a Climate Data Record

Presentation type: Type Poster

Recent years have seen the fast development of data mining approach and the subsequent methods to extract meaningful information from large and heterogeneous datasets. From predicting stock values to recognizing your friends’ faces in your photo album, machine learning/data mining algorithms are successfully applied in a very wide range of applications. These methods take advantage of massive multivariate datasets to perform accurate predictions. Altimetry databases do fit in this category.

At the moment, SSH data editing along the nadir track is performed using simple statistical techniques based on empirical and theoretical expertise about the observed quantities (thresholding, variance checks). This method, tailored to the specificity of each instrument, has demonstrated its ability to provide high quality data, but, in the perspective of future altimetry missions (as SWOT) and the revolution of 2D SSH measurements, a more robust and more generic approach is needed.

The purpose of this study is to experiment with classification methods such as neural networks, random forests, and logistic classification… to assess their potential for the editing of satellite altimetry data. We present results from an exploratory analysis of different case studies: ice flagging on SARAL/AltiKa data and rain cells editing in simulated SWOT scenes. We examine the performances of different methods and their sensitivity. This family of methods might become very useful for the validation of SWOT data, which will share many aspects with image processing.


Poster show times:

RoomStart DateEnd Date
Grande Halle Thu, Nov 03 2016,11:00 Thu, Nov 03 2016,18:00
Pierre Prandi