Introduction to the FIPPI Algorithm
The Fast Iterative Pixel Purity Index (FIPPI) is an effective method for extracting endmember signatures from hyperspectral remote-sensing data. It is an iterative improvement of the pixel purity index (PPI), designed to improve computational efficiency and obtain better endmember estimates through a sound stopping condition and carefully designed iteration rules.
Mathematical Principles of the Algorithm
Dimensionality Reduction
In the FIPPI algorithm, the original high-dimensional data cube is first reduced in dimension in order to lower computational complexity. Common dimensionality-reduction methods include:
Principal component analysis (PCA): project the high-dimensional data into principal-component space and retain the principal information.
Maximum noise fraction (MNF) transform: order the components by image quality and retain the high-quality components.
For the original data matrix \(\boldsymbol{X} \in \mathbb{R}^{m \times n}\), dimensionality reduction produces \(\boldsymbol{X}^{*} \in \mathbb{R}^{m \times p}\), where \(p < n\).
Automatic Target Generation Process (ATGP)
The ATGP algorithm estimates the initial endmember set from an orthogonal subspace projection. Given the dimensionality-reduced data
\(\boldsymbol{X}^{*}\), which contains \(p\) feature vectors, the ATGP algorithm is as follows:
- Find the index \(i_1\) of the first endmember vector that maximizes
\(\sum_{j=1}^{p} (\boldsymbol{x}_j^*)^2\), that is, \(\boldsymbol{u}_1 = \boldsymbol{x}_{i_1}^*\).
- For the \(k\)th (\(k \geq 2\))
endmember vector, solve:
\[ \boldsymbol{u}_k = \arg\max_{\boldsymbol{x}_j^*} \left\|\left(\boldsymbol{I}_p - \boldsymbol{U}_{k-1}\boldsymbol{U}_{k-1}^+\right)\boldsymbol{x}_j^*\right\|^2 \]where
\[ \boldsymbol{U}_{k-1} = \begin{bmatrix}\boldsymbol{u}_1 & \boldsymbol{u}_2 & \cdots & \boldsymbol{u}_{k-1}\end{bmatrix} \], and \(\boldsymbol{U}_{k-1}^+\) is the pseudoinverse of \(\boldsymbol{U}_{k-1}\).
The algorithm repeats the process above until \(p\)
initial endmember vectors are obtained.
FIPPI Iteration Rules
After the initial endmember set \(\boldsymbol{U} \in \mathbb{R}^{p \times p}\) is obtained with the ATGP algorithm, FIPPI refines it according to the following iteration rules:
Project the dimensionality-reduced data onto the current endmember space: \(\boldsymbol{Y} = \boldsymbol{U}^T\boldsymbol{X}^*\)
For each sample, find its maximum absolute projection value in the projected space and the corresponding sample index
\(i_j\):
\[ i_j = \arg\max_{1 \leq i \leq m} \left|\boldsymbol{y}_i\right| \]Construct the extrema set \(\boldsymbol{X}^{ext} = \left\{\boldsymbol{x}_{i_j}^*\right\}_{j=1}^q\), where \(q\) is the number of extrema.
Check the stopping condition: if \(\boldsymbol{X}^{ext} \subseteq \boldsymbol{U}\), stop the iteration; otherwise, add to \(\boldsymbol{U}\) the vectors in \(\boldsymbol{X}^{ext}\) that are not already in \(\boldsymbol{U}\), and return to step 1 to continue iterating.
After several iterations, the algorithm outputs \(\boldsymbol{U}\) as the estimated endmember set.
MATLAB Implementation
The following is a MATLAB implementation of the FIPPI algorithm. See the code comments for details:
| |
Usage example:
| |
Summary
Through a sound mathematical model and carefully designed iteration rules, the FIPPI algorithm effectively extracts endmember signatures from hyperspectral remote-sensing data. It overcomes several shortcomings of the conventional PPI algorithm, improves computational efficiency and endmember estimation accuracy, and is a powerful tool for endmember extraction in remote sensing.
