| DOI | Resolve DOI: https://doi.org/10.1109/SSCI51031.2022.10022301 |
|---|
| Author | Search for: Valdes, Julio J.1; Search for: Pou, Antonio |
|---|
| Affiliation | - National Research Council of Canada. Digital Technologies
|
|---|
| Format | Text, Article |
|---|
| Conference | 2022 IEEE Symposium Series on Computational Intelligence (IEEE SSCI), Dec. 4-7, 2022, Singapore |
|---|
| Subject | explainable AI; climate change; atmospheric patterns; water vapor images; convolutional neural networks |
|---|
| Abstract | Large stocks of meteorological satellite images contain important climatic information of the last decades. This paper explores the capabilities of three Explainable AI procedures (Occlusion, XRAI and SHAP) as tools to analyze and uncover patterns in the mid to upper troposphere Water Vapor dynamics, and their relationships with yearly seasons. The data consisted of 4140 daily Meteosat satellite images on the WV6.2 band between 2010 to 2020 and the base model was a convolutional neural network targeting season prediction. The results show that each explanation procedure highlights different aspects of the atmospheric processes and are appropriate for different purposes: Occlusion for studying the general traits of the 11 years time period covered by the images, XRAI for detecting and following dynamic patterns for short spans of time, and SHAP for detailed geographical expressions of atmospheric processes. All together Explainable AI exhibits an important potential for the study of atmospheric and climate change, which should be further investigated. |
|---|
| Publication date | 2023-01-30 |
|---|
| Publisher | IEEE |
|---|
| In | |
|---|
| Language | English |
|---|
| Peer reviewed | Yes |
|---|
| Export citation | Export as RIS |
|---|
| Report a correction | Report a correction (opens in a new tab) |
|---|
| Record identifier | f8db6bc2-d9c9-4b34-be3f-586b6d331b71 |
|---|
| Record created | 2023-01-31 |
|---|
| Record modified | 2023-02-01 |
|---|
