Investigation of meso-scale Sentinel-3 product along-track correlations and the potential of inter-track SSHA estimation using machine learning

Abstract

Satellite altimetry is an important technology used to measure sea level with high spatial and temporal resolution. Sentinel-3, a Copernicus satellite mission, offers three types of variables captured simultaneously for the first time; sea level (SSH), sea surface temperature (SST) and ocean colour (OC) variables. Sea level is measured with SAR altimetry, a technique that considerably increases spatial resolution compared to other means of observation. Altimetry measures sea level across a line that coincides with the satellite ground track, whereas SST and OC are measured on a grid. What we lack are sea level observations in-between ground tracks that would better resolve meso-scale variability. This thesis is focused on two objectives, considering previous work that has indicated associations between those variables. The first objective was to investigate the correlations among SSH, SST and OC, while the second objective was to assess to what extent inter-track sea level can be estimated using SST and OC as predictors in machine learning algorithms. Daily Sentinel-3 data over a period of eleven months were pre-processed and brought into a form that allowed for computation of metrics such as auto- and cross-correlations in the along-track direction. The focus was on the spatial scales that would enable to detect meso-scale features, such as eddies. With respect to the inter-track sea level estimation two paths were followed. In the first path, Random Forest (RF) and Multilayer Perceptron (MLP) were applied using the complete 11-month dataset as input. Moreover, RF was applied on input data that belong to each separate day. In the second path, 1D Convolutional Neural Network (CNN) was used on the complete 11-month dataset, which inherently honors the spatial dependency of the variables in contrast to the first path. Generally, the correlations between the variables were found to exist in the meso-scale but were not always strong and they depend on several other factors, such as meteorological conditions, scales included in the analysis and techniques used. All three techniques -RF, MLP and 1D CNN- that were applied on the complete 11-month dataset gave poor results. On the contrary, when RF was applied on the per-day data gave promising results that are reliable mostly in the vicinity of the ground track, although they are not based on one single global model. The results from this project suggest that there must be more research on the correlation analysis of Sentinel-3 data. It can be improved by using additional or similar techniques, such as localized cross-correlation metrics on various spatial scales. With respect to the inter-track sea level estimation, far more investigation is needed. However, there are indications that a machine learning data-driven approach could potentially work to some extent. Sentinel-3 data will become more abundant in the next years which will assist data science algorithms such as CNNs which require huge datasets.