Sparse Unmixing of Hyperspectral Data
read more
Citations
Hyperspectral Unmixing Overview: Geometrical, Statistical, and Sparse Regression-Based Approaches
Hyperspectral Unmixing Overview: Geometrical, Statistical, and Sparse Regression-Based Approaches
Hyperspectral Remote Sensing Data Analysis and Future Challenges
Spectral mixture modeling - A new analysis of rock and soil types at the Viking Lander 1 site. [on Mars]
Spectral unmixing
References
Atomic Decomposition by Basis Pursuit
Signal Recovery From Random Measurements Via Orthogonal Matching Pursuit
Signal Recovery from Random Measurements Via Orthogonal Matching Pursuit: The Gaussian Case
Decoding by linear programming
Stable signal recovery from incomplete and inaccurate measurements
Related Papers (5)
Vertex component analysis: a fast algorithm to unmix hyperspectral data
Frequently Asked Questions (14)
Q2. What are the future works in "Sparse unmixing of hyperspectral data" ?
Although their experimental results ( conducted with both simulated and real data sets ) are very encouraging, there are several aspects to be considered in practice, and they are worthy to be further investigated in future research efforts. The final issue that is to be explored in future developments is the high computational complexity of the sparse unmixing algorithms, addressed in this paper by the consideration of the fast algorithms based on the augmented Lagrangian method of multipliers, but they are also subject to further improvements related to the inherently parallel nature of such algorithms. This feature anticipates the high scalability of the potential parallel solutions to this approach. To conclude this section, the authors would like to emphasize their significant efforts in testing the most suitable parameters in order to report only the near-optimal results for each considered method.
Q3. What is the main rationale for using this threshold?
The main rationale for using this threshold is that, after inspecting the results of different unmixing scenarios, the authors concluded that a reconstruction attaining SRE(dB) = 5 dB is still useful.
Q4. What was the process used to extract the endmembers from the simulated data?
two endmember extraction algorithms (VCA and N-FINDR) were used to automatically extract the endmembers from the simulated data.
Q5. What is the main issue in the evaluation of the sparse unmixing algorithms?
An important issue in the evaluation of the sparse unmixing algorithms is their computational complexity, particularly when large spectral libraries are used to solve the unmixing problem.
Q6. What was the spectral library used in the sparse unmixing process?
In this library, only materials with a spectral angle of at least 3◦ with regard to other materials in the library were retained in order to avoid strong similarities between the spectral signatures when conducting the sparse unmixing process.
Q7. What is the condition for a s-sparse vector solution of (P?
As with problem (P0), the condition γ2s < 4 √ 2− 3 2.6569 referred to in Section II-A1 is now applied to the restricted isometric constants of matrix D to ensure that any s-sparse vector solution of (P+0 ) is recovered by solving the problem (P + 1 ).
Q8. What is the way to solve the sparse unmixing problem?
It is important to emphasize that, by setting λ = 0 in (24), one can arrive to an LS solution of the system, which is obtained by solving the unconstrained optimization problem(PLS) : min x ‖y −Ax‖2. (25)The solution of optimization problem (25) has a poor behavior in terms of accuracy when the matrix of coefficients is ill conditioned (as it is always the case in the sparse unmixing problem, in which the authors deal with fat matrices) or when the observations are affected by noise.
Q9. How did the authors generate the simulated hyperspectral image?
the authors randomly selected five of the spectral signatures in the resulting subset and used them to generate a simulated hyperspectral image with 75 × 75 pixels and 224 bands per pixel.
Q10. How does the algorithm find an optimal endmember set?
It finds an optimal endmember set by examining the change in the root-mean-square error (rmse) after reconstructing the original scene using the fractional abundance estimations, as shown in Algorithm 2.
Q11. What is the interesting case study?
From the viewpoint of their considered problem, perhaps, this is the most interesting case study since the noise in the hyperspectral images is usually correlated.
Q12. What is the difference between the DCT of the columns of A and the L L matrix?
This results from the following.1) Computing the DCT of the columns of A is equivalent to left multiplying A by a unitary L× L matrix, which does not therefore change spark(A).
Q13. What is the general trend of the algorithm behavior in the simulated scenario?
In general, the algorithm behavior observed in previous simulated scenarios is confirmed here, with the general trend that most considered approaches perform better in the presence of correlated noise rather than in the presence of white noise.
Q14. What is the probability of success for all methods when the cardinality is higher than ten?
For the libraries composed of real signatures (A1, . . . ,A4), the probability of success is low for all methods when the cardinality is higher than ten.