Statistical Learning in Earth Monitoring

Event details
Date | 23.10.2012 |
Hour | 16:15 › 17:15 |
Speaker | Professor Gustavo Camps-Valls, Dept. of Electronics Engineering, Valencia University, Spain |
Location | |
Category | Conferences - Seminars |
The last few hundred years, human activities have precipitated an environmental crisis on Earth. Quantification of the impact on the Earth's system are matter of current and intense research. In the last decade, advanced statistical methods have been introduced to help monitoring our impact on the vegetation and atmosphere. This seminar will concentrate on two recent efforts conducted in my lab to constrain the problem by using remote sensing data and advanced statistical inference models.
Solar-induced chlorophyll fluorescence (ChF) provides one of the most innovative sources of information to monitor plant activity from space. Accurate and robust estimation of leaf chlorophyll content is needed to calculate the ChF yield and to allow extrapolation at the landscape scale. Vegetation indices typically become either too general or specific to particular scenarios, and they under-exploit the wealth of the spectral information. Conversely, physically-based methods such as inverting a radiative transfer (RT) model, require site-specific information for proper model parametrization, which is not always available. Alternatively, non-parametric models provide generally higher accuracies, and recent statistical methods like Gaussian processes also provide information about the uncertainty of the estimation, which may lead to improved quantification of gross primary productivity (GPP). Applications in airborne data will illustrate the capabilities of the methods, in preparation of the upcoming ESA Fluorescence Explorer (FLEX) mission.
The second case study will focus on the estimation of temperature, water vapor, and ozone concentration from spaceborne very high spectral resolution infrared sounding instruments. These sensors can be used to calculate the profiles of such atmospheric parameters with unprecedented accuracy and vertical resolution. Despite the constant advances in sensor design and retrieval techniques, it is not trivial to invert the full information of the atmospheric state contained by such hyperspectral measurements. Numerical inversion of an RT model gives accurate results but shows numerical instability and high computationally burden. Advantageously, statistical regression methods enable fast retrievals from high volumes of data, rank the most relevant input variables and observations, and provide confidence intervals for the predictions. Evidence of the advantages given by such approaches in the upcoming Meteosat Third Generation Infrared Sounder (MTG-IRS) data will be given.
Solar-induced chlorophyll fluorescence (ChF) provides one of the most innovative sources of information to monitor plant activity from space. Accurate and robust estimation of leaf chlorophyll content is needed to calculate the ChF yield and to allow extrapolation at the landscape scale. Vegetation indices typically become either too general or specific to particular scenarios, and they under-exploit the wealth of the spectral information. Conversely, physically-based methods such as inverting a radiative transfer (RT) model, require site-specific information for proper model parametrization, which is not always available. Alternatively, non-parametric models provide generally higher accuracies, and recent statistical methods like Gaussian processes also provide information about the uncertainty of the estimation, which may lead to improved quantification of gross primary productivity (GPP). Applications in airborne data will illustrate the capabilities of the methods, in preparation of the upcoming ESA Fluorescence Explorer (FLEX) mission.
The second case study will focus on the estimation of temperature, water vapor, and ozone concentration from spaceborne very high spectral resolution infrared sounding instruments. These sensors can be used to calculate the profiles of such atmospheric parameters with unprecedented accuracy and vertical resolution. Despite the constant advances in sensor design and retrieval techniques, it is not trivial to invert the full information of the atmospheric state contained by such hyperspectral measurements. Numerical inversion of an RT model gives accurate results but shows numerical instability and high computationally burden. Advantageously, statistical regression methods enable fast retrievals from high volumes of data, rank the most relevant input variables and observations, and provide confidence intervals for the predictions. Evidence of the advantages given by such approaches in the upcoming Meteosat Third Generation Infrared Sounder (MTG-IRS) data will be given.
Practical information
- General public
- Free
- This event is internal
Organizer
- IIE
Contact
- Dr Tuia Devis and Professor François Golay, LASIG