Find the EGU on

Tag your tweets with #EGU16

Please note that this session was withdrawn and is no longer available in the respective programme. This withdrawal might have been the result of a merge with another session.

NP1.4

Machine learning adaptations for Earth monitoring
Convener: Nicolas Brodu  | Co-Convener: Frederic Frappart 
Sensors on-board satellites combines many sources of information, like multispectral and radar images, microwave brightness temperatures and emissivities, radar scatterometry, etc. When combined with in situ measurements, properly exploiting this wide variety of data sources, of heterogeneous temporal and spatial resolutions, become a challenge. Moreover, typical applications like land use occupation or changes detection caused by natural and anthropogenic phenomenon (e.g., flood, fires, forest logging), rely on automated inference at some point. Although generic Machine Learning (ML) techniques are well-established, their general framework is not necessary relevant for environmental applications. For example, how to deal with scales ranging from local measurements to the swath of the satellite ; accounting for the spatial consistency between nearby samples in ML algorithms ; or including the known seasonal and interannual variations to build better time series descriptors. Similarly, the physics of observed phenomena need to be incorporated into ML frameworks for accurate modeling and predictions. Therefore, this session calls for abstracts on how to best apply ML methods in Earth Science contexts. The goal is to promote exchanges and ideas on ML techniques that could benefit multiple Earth Science domains and questions of interest ; domain-specific ML tricks that focus on only one particular Earth Science application may be interesting to other EGU sessions, but are here out of topic. Examples of relevant abstracts include new methodologies for data fusion (e.g. SAR/multispectral/in situ measurements), how to include physics of natural processes in
feature descriptors (e.g. fluid dynamics), how ML can help interpret large data sets (e.g. automated detection of areas with different statistical properties), how multi-resolution (spatial or temporal) features help detect and deal with characteristic scales, how new features can be extracted from data sets using innovative image processing and time series analysis, etc.