Analysis of spatial scales in satellite data reconstructed using a neural network
|Analysis of spatial scales in satellite data reconstructed using a neural network
|Translated title :
|[fr] Analyse des échelles spatiales dans des données satellitaires reconstruites à l’aide d’un réseau neuronal
|Date of defense :
|Alvera Azcarate, Aida
|Committee's member(s) :
|Life sciences > Aquatic sciences & oceanology
|Université de Liège, Liège, Belgique
|Master en océanographie, à finalité approfondie
|Master thesis of the Faculté des Sciences
[en] Sea surface temperature (SST) is a very important variable to assess numerous physical and biological phenomena, most notably the ocean’s impact on the recent climatic changes. It’s also a parameter that can be measured regularly and over the entire surface of the oceans and seas thanks to satellites. The major issue with those measurements are data breaches due to clouds. Numerous softwares can approximate the lacking data with various methods, one of those being the use of a neural network with a U-Net architecture (Siddique et al, 2021) named DINCAE (Data-Interpolating Convolutional Auto-Encoder) (Barth et al., 2020 et 2022).
This report aims to improve the reliability of the reconstructions made by DINCAE at differing scales, most notably small scales, with 5 or 6 neural layers. To achieve this, it will focus on the use of skip-connections between the input and the output of different layers. The reliability assessment will be done in 2 ways. +The first one is to evaluate the RMS error of the reconstructed data compared to the initial data, including purposefully removed data then restored. The second one is to assess the amount of variance kept by the software as reconstruction softwares are known to under-estimate it. This assessment is done annually and seasonally. The study zone chosen for this report is the Alboran sea, the most western part of the Mediterranean sea.
The work’s result shows the most optimal skip-connections to use are dependent on the scales and period studied. No configuration could be found to both minimize the RMS error as well as to maximize the variance, we thus need to find a trade-off between the two. However, a 6 layers configuration with a skip-connection on each of the first 5 layers has been found to minimize the RMS error, while a 6 layers configuration with a skip-connection only on the second layer has been found to keep the most variance in the system over the entire year at most scales, although it’s not the best over each and every one of the season.
Cite this master thesis
The University of Liège does not guarantee the scientific quality of these students' works or the accuracy of all the information they contain.