Igarss 2016

From [2016-07-10 Sun] to [2016-07-15 Fri], I attend the 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2016) in Beijing, China. As usual, meeting colleagues and friends is a good way to close an intense academic year. The quality of the conference was good, and I found a lot of interesting papers. I think we should deeply consider convolutional neural nets, it seems to be trendy …

Maïlys Lopes, a PhD student working with me and Stéphane Girard, has presented her work on Grassland monitoring using time series: http://igarss2016.org/Papers/viewpapers.asp?papernum=3617.

For your information, please find the paper I have found interesting for my research purpose (which is a strong filter!):

[2016-07-11 Mon]

“CLODD BASED BAND GROUP SELECTION”
  • Do bands grouping
  • Ref to a band selection method using proximity measure (ref 5)
  • Does not talk about the extracted feature :(
14:50: “OPTIMIZING REMOTE SENSING OF NATIVE GRASSLANDS FOR MONITORING AND CHANGE DETECTION”, Room 3011+B
  • Introduction sur pourquoi il est important de gérer les “grasslands”
  • Actuellement à la main, c’est pas possible de le faire sur long terme
  • Utilisation de l’optique et radar en même temps

  • Questions:

    • best combination of optical and radar ?
    • quels sont les données qui couvrent correctment le territoire ?
  • N’utilise pas des SITS, mais un couple Radarsat-2/Landsat-5.
  • Différencier les prairies des cultures ?
  • Classif pixel par SVM, avec differents combination of inputs
15:40: “Synergy of Sentinel-1 and Sentinel-2 imagery for wetland monitoring …”
  • In order to use the images stream …
Poster “A COMPREHENSIVE EVALUATION OF FEATURE SELECTION ALGORITHMS IN HYPERSPECTRAL IMAGE CLASSIFICATION”
  • Simple FS methods are tested
  • Forward strategy is uses with SVM -> very long of course.
  • test: stabilty, accuracy and processing time.
  • Methods that show high accuracy are not stable and vise versa.

[2016-07-12 Tue]

09:00: “FAST HYPERSPECTRAL IMAGE DENOISING BASED ON LOW RANK AND SPARSE REPRESENTATIONS”, Room3021+B
  • Use a subspace model, which is similar to HDDA
  • Provide an algorithm to remove noise from HSI
09:20: “AN APPROACH TO CONIFER SPECIES CLASSIFICATION BASED ON CROWN STRUCTURE MODELING IN HIGH DENSITY AIRBORNE LIDAR DATA” , Room 311B
  • Provide a model to estimate branches of trees based on PCA
  • Derive crown indices from the PCA modeling
  • Do classification Sparse SVM (?)
09:40: “LIDAR INFORMATION EXTRACTION BY ATTRIBUTE FILTERS WITH PARTIAL RECONSTRUCTION”, Room 311B
  • Partial reconstruction is used to filter LiDar data
  • EMP is constructed and used as input to classifier
16:00: “TREE HEIGHT ESTIMATES IN BOREAL FOREST USING GAUSSIAN PROCESS REGRESSION”, Room 301A+B

[2016-07-13 Wed]

IDEAL REGULARIZED KERNEL FOR HYPERSPECTRAL IMAGE CLASSIFICATION
  • Work quite well: enlarge the area where the hyperparameter of the kernels works well
  • Not clear how to fit the regulatization parameter.
A GAUSSIAN APPROACH TO SUBSPACE BASED CLASSIFICATION OF HYPERSPECTRAL IMAGES
  • Each class is modeled by a mixture of Gaussian distribution
  • The Gaussian distribution is modeled by random projection
GETTING PIXELS AND REGIONS TO AGREE WITH CONDITIONAL RANDOM FIELDS
  • Propose a model to share contextual information between pixel-classifier and region-classifier

  • Spatial contrains:

    • inside a region: penalize if a pixel inside a region is of a different label than the region
    • between pixel: potts model
    • between regions: potts model between region
FOREST BIODIVERSITY MAPPING USING AIRBORNE LIDAR AND HYPERSPECTRAL DATA

[2016-07-14 Thu]

DETECTION OF COMPOUND STRUCTURES BY REGION GROUP SELECTION FROM HIERARCHICAL SEGMENTATIONS
  • Propose a method to detect “compound structure” is urban area
  • A markov random field on spatial primitive is proposed
  • Gibbs sampler is used to optimize the field
TRANSFORMING THE AUTOCORRELATION FUNCTION OF A TIME SERIES TO DETECT LAND COVER CHANGE
  • Detect changes by looking stationarity in the autocorrelation function
  • window filter is used when the autocorrelation is computed to remove edges effect. Better results are observed.

[2016-07-15 Fri]

A DATA-DRIVEN IDENTIFICATION OF GROWTH-MODEL CLASSES FOR THE ADAPTIVE ESTIMATION OF SINGLE-TREE STEM DIAMETER IN LIDAR DATA
  • high intensity lidar for individual tree caracterization

  • extraction of several parameters modeling

    • the crown structure
    • the forest density
    • the topography
SPARSE DISTRIBUTED HYPERSPECTRAL UNMIXING
  • unmixing of hyperspectral image with ADMM
  • can be used on large data set
  • but the solution is not identical when the method is distributed

Comments

comments powered by Disqus