Accurately and promptly detecting changes in land-surface features is essential for environmental monitoring and resource management. Compared with conventional single-band or multispectral remote sensing imagery, hyperspectral imagery contains a large amount of spectral information and therefore performs better at change detection. This article introduces a subspace-based change detection (SCD) method for hyperspectral images.
Subspace Analysis for Hyperspectral Imagery
Subspace learning is an effective tool for analyzing hyperspectral imagery. Its strengths are that it can accurately describe linearly mixed signals and that it offers better signal separability. Orthogonal subspace projection (OSP) is a successful method for hyperspectral image analysis and is widely used across a range of applications.
A mixed pixel in a hyperspectral image can be described by the following linear model:
\[ \mathbf{r} = \alpha\mathbf{s} + \sum_{i=1}^{p}\beta_i\mathbf{v}_i + \mathbf{n} \]where \(\mathbf{r}\) is the spectral signature of the mixed pixel, \(\mathbf{s}\) is the target spectral signature,
\[ \mathbf{V} = \begin{bmatrix}\mathbf{v}_1 & \mathbf{v}_2 & \cdots & \mathbf{v}_p\end{bmatrix} \]is the matrix of non-target spectral signatures, \(\alpha\) and \(\beta_i\) are the abundances of the corresponding signatures, and \(\mathbf{n}\) is a random noise vector.
The OSP operator can be written as
\[ \mathbf{P}_{\mathbf{V}^{\perp}} = \mathbf{I} - \mathbf{V}(\mathbf{V}^T\mathbf{V})^{-1}\mathbf{V}^T \]where \(\mathbf{P}_{\mathbf{V}^{\perp}}\) is the orthogonal-complement projector that maps a feature vector onto the orthogonal complement of \(\mathbf{V}\). If the spectral signatures of all distinct materials are known in advance, OSP can suppress the non-target signatures and assess whether the target signature is present.
Subspace-Based Change Detection (SCD)
In change detection, the pixel observed at time 2 is treated as a potential change target, and the background subspace is built from the spectral signature of the corresponding pixel at time 1 together with additional spectral or spatial information. The difference between the target and the background is then measured to decide whether a genuine change is present. Subspace distance is used to quantify the spectral change.
If the observed pixel does not change over the time interval, its spectrum lies almost entirely inside the background subspace. That subspace is constructed from the unchanged corresponding pixel, which has a spectral signature similar to the observation. A pixel that has changed therefore shows a relatively large subspace distance from the background subspace, whereas an unchanged pixel yields a distance close to zero.
Let \(\mathcal{U}\) be a Euclidean space that can be written as the direct sum of a subspace \(\mathcal{S}\) and its complement \(\mathcal{S}^{\perp}\), that is, \(\mathcal{U} = \mathcal{S} \oplus \mathcal{S}^{\perp}\). For any vector \(\mathbf{x} \in \mathcal{U}\), we have \(\mathbf{x} = \mathbf{x}_{\mathcal{S}} + \mathbf{x}_{\mathcal{S}^{\perp}}\), where \(\mathbf{x}_{\mathcal{S}} \in \mathcal{S}\) and \(\mathbf{x}_{\mathcal{S}^{\perp}} \in \mathcal{S}^{\perp}\). If every vector \(\mathbf{x}_{\mathcal{S}^{\perp}}\) is orthogonal to \(\mathcal{S}\), then \(\mathcal{S}^{\perp}\) is called the orthogonal complement of \(\mathcal{S}\).
The operator that projects a vector \(\mathbf{x}\) onto the subspace \(\mathcal{S}\) is
\[ \mathbf{P}_{\mathcal{S}} = \mathbf{B}(\mathbf{B}^T\mathbf{B})^{-1}\mathbf{B}^T \]where
\[ \mathbf{B} = \begin{bmatrix}\mathbf{b}_1 & \mathbf{b}_2 & \cdots & \mathbf{b}_p\end{bmatrix} \]is the matrix whose columns are the basis vectors of \(\mathcal{S}\).
The subspace distance from \(\mathbf{x}\) to \(\mathcal{S}\) equals the norm of the projection of \(\mathbf{x}\) onto \(\mathcal{S}^{\perp}\). Using the projector above, the projection of \(\mathbf{x}\) onto \(\mathcal{S}^{\perp}\) is
\[ \mathbf{x}_{\mathcal{S}^{\perp}} = (\mathbf{I} - \mathbf{P}_{\mathcal{S}})\mathbf{x} = (\mathbf{I} - \mathbf{B}(\mathbf{B}^T\mathbf{B})^{-1}\mathbf{B}^T)\mathbf{x} \]The subspace distance is the norm of \(\mathbf{x}_{\mathcal{S}^{\perp}}\):
\[ d(\mathbf{x}, \mathcal{S}) = \Vert\mathbf{x}_{\mathcal{S}^{\perp}}\Vert \]To simplify the computation, define \(\mathbf{y} = (\mathbf{B}^T\mathbf{B})^{-1}\mathbf{B}^T\mathbf{x}\), so that \(\mathbf{x}_{\mathcal{S}} = \mathbf{B}\mathbf{y}\) and \(\mathbf{x}_{\mathcal{S}^{\perp}} = \mathbf{x} - \mathbf{B}\mathbf{y}\). The simplified operator is then equivalent to the original one:
\[ d(\mathbf{x}, \mathcal{S}) = \Vert\mathbf{x} - \mathbf{B}\mathbf{y}\Vert \]To remove the effect of spectral magnitude on the subspace distance, the spectral vectors are normalized. The SCD operator can finally be written as
\[ \text{SCD}(\mathbf{x}) = \frac{\Vert\mathbf{x} - \mathbf{P}_{\mathbf{B}}\mathbf{x}\Vert}{\Vert\mathbf{x}\Vert} \]where \(\mathbf{B}\) is the background subspace formed from the spectral signature of the corresponding pixel at time 1 and from additional information. This operator computes the subspace distance from the unit spectral vector at time 2 to the background subspace, and thereby measures the spectral change over the time interval.
Constructing the Background Subspace
The background subspace is formed from the spectral signature of the corresponding pixel and from additional information. A pixel pair is the basic unit for detecting spectral change, while the additional information is optional and is chosen for a particular application. Three ways of constructing the background subspace are given below.
- Construction from spectral information
Spectral signatures of non-target land-cover types are added to the background subspace so that unwanted changes are suppressed. The background subspace \(\mathcal{A}\) can be written as
\[ \mathcal{A} = \text{span}\{\mathbf{r}_c, \mathbf{v}_1, \mathbf{v}_2, \cdots, \mathbf{v}_p\} \]where \(\mathbf{r}_c\) is the spectral signature of the corresponding pixel and \(\mathbf{v}_i\) is the spectral signature of a target class.
- Construction from spatial information
Spatial information can reduce false detections caused by misregistration. Two ways of using spatial information are proposed:
- Local SCD (LSCD): center a subwindow on the corresponding pixel and build the background subspace from every pixel inside that window:
where \(\mathcal{W}\) is the set of all pixels in the subwindow.
- Adaptive SCD (ASCD): for each pixel \(\mathbf{r}\), search the subwindow centered on its corresponding pixel and select the pixel that gives the smallest SCD value with \(\mathbf{r}\). Use that pixel to build the background subspace:
By exploiting spatial information, SCD can reduce false detections caused by misregistration. When the images are misregistered, the true corresponding pixel still lies inside the subwindow, so the correct spectral change can still be measured.
The subwindow size depends on the degree of misregistration. In general, the window should be large enough to cover the magnitude of the misregistration, but an overly large window increases the computational cost and may miss genuine changes. Besides square windows, other shapes can be considered, such as circular or asymmetric windows.
Both spectral information and spatial information are additional inputs to SCD. When spatial information is used to improve SCD, spectral information can be included at the same time to suppress unwanted changes. The background subspace can be constructed flexibly, and other kinds of additional information—such as a wider range of spectral signatures—can be explored so that the algorithm can be adapted to different applications.
Summary
Overall, SCD measures changes in spectral signatures directly through subspace learning. It makes full use of the rich spectral information in hyperspectral imagery and detects changes more effectively than the conventional per-band differencing method. The background subspace can be constructed in many flexible ways and can be refined and optimized for different applications.
