CICYT TIN2015-64210-R

WP1 Improvement of algorithms

  • WP1.1 Structured multiscale and multiresolution domains.
    • EO data are related, structured data (spectral, spatial and time).
    • Kernel methods (GPs and DNNs) to account for spectral, spatial and time structures.
    • Develop:
      1. Tensorial convolutional nets in space-spectrum domains (for optical and infrared sensors).
      2. Spatio-temporal FIR deep nets (alternative to convolutional nets + long short-term memory units).
      3. Structured deep kernel regression without pre-images (i.e., Structured KRR).
  • WP1.2 Multivariate outputs.
    • Study the field of 'structured output learning', related to 'multitask learning'.
    • Projects: 1) extend GPs, 2) orthonormalized kernel feature extraction methods.
  • WP1.3 Uncertainty estimation and propagation.
    • Derive tighter predictive variances for GPs.
      1. coarse-to-fine unsupervised covariances in GPs.
      2. combination of experts via standard averaging or geometric distance weighting.
    • A direct measure of the uncertainty is the Jacobian of the transformation.
      • A high determinant of the Jacobian means small changes in inputs affect predictions drastically.
  • WP1.4 Including physical knowledge via emulation of physical RTMs.
    • Use emulators to approximate physical models ⇒ Develop a full toolbox of emulators.
      • PROSAIL (done!), OSS (Optimal Spectral Sampling) for atmospheric apps., MODTRAN, etc.

WP2 Efficient implementations

  • WP2.1 Deal with kernel and deep architectures for EO big data ⇒ reduce the kernel size using approximations.
  • WP2.2 Running/adapting algorithms for multicore CPUs and GPUs.
  • WP2.3 Divide and conquer strategies for RS data.

WP3 Extracting information from models

  • WP3.1 Feature ranking and global sensitivity analysis (GSA).
    1. Develop GSA techniques for the predictive mean and variance of GPs for several covariances functions.
    2. Introduce kernel version of standard methods for GSA based on estimating Euclidean distances in input spaces.
    3. Analyze the sensitivity scores for physical RTM emulators and its optimization.
  • WP3.2 Inspecting deep features.
    • DNNs ⇒ inspecting the learned transformation at each layer ⇒ gives the relationships between input variables.
  • WP3.3 Causality.

WP4 Applications for RS and geosciences

  • WP4.1 (Vegetation) Sentinel 2/3 data processing through RTM emulation.
  • WP4.2 (Atmosphere) Estimation of atmospheric profiles with super-spectral infrared sounders.
  • WP4.3 (Carbon/Heat) Estimation of global time-resolved carbon and head fluxes.

WP5 Management and transfer

  • WP5.1 Reports. Publications in international conferences and journals.
  • WP5.2 Toolboxes and databases.
  • WP5.3 Special sessions/workshops.