Mixture Tuned Matched Filtering (MTMF) is an advanced target detection algorithm designed for hyperspectral and high–spectral–resolution imagery. It combines a noise-whitened matched filter with a physically constrained infeasibility metric to reduce false positives.

Unlike simple spectral similarity metrics, MTMF accounts for background statistics and mixture plausibility, making it particularly robust for mineral detection and subtle material mapping.

How it Works #

MTMF is composed of four main stages:

1. Background statistics estimation
2. Noise whitening
3. Minimum Noise Fraction (MNF) transform
4. Matched filtering + mixture tuning (infeasibility)

Each step is computed explicitly over the Area of Interest (AOI).

1. Data Representation #

Let each pixel spectrum be defined as: $R = [R₁, R₂, …, Rₙ]ᵀ$

Let the target (reference endmember) spectrum be: $S = [S₁, S₂, …, Sₙ]ᵀ$

where n is the number of spectral bands

2. Background Mean and Noise Estimation #

2.1 Mean-centering #

The mean spectrum over the AOI is computed: $μ = E[R]$

Each pixel is mean-centered: $X = R − μ$

2.2 Noise Estimation (Shift-Difference Method) #

Noise is estimated spatially using a shift-difference operator:

$N(i,j) = ( R(i,j) − 0.5·R(i−1,j) − 0.5·R(i,j−1) ) / √1.5$

This approximates high-frequency noise while preserving large-scale background structure.

The noise covariance matrix is then: $Σₙ = Cov(N)$

Regularization is applied: $Σₙ ← Σₙ + εI$

with $ε = 10⁻⁶$ for numerical stability.

3. Noise Whitening #

Eigen-decomposition of the noise covariance: $Σₙ = Eₙ Λₙ Eₙᵀ$

Whitening matrix: $W = Λₙ⁻¹ᐟ² Eₙ$

Noise-whitened data: $Xₙw = W X$

After whitening, noise variance is normalized to approximately 1 in all directions.

4. Minimum Noise Fraction (MNF) Transform #

The covariance of the whitened data is computed: $Σₙw = Cov(Xₙw)$

Eigen-decomposition: $Σₙw = Eₘₙf Λ_bg Eₘₙfᵀ$

The MNF transform is: $Xₘₙf = Eₘₙf W X$

In the implementation, the whitening and MNF rotations are fused:

$T_full = Eₘₙf W$

$Xₘₙf = T_full X$

This reduces per-pixel matrix multiplications and improves performance in large AOIs.

The target spectrum is transformed identically: $Sₘₙf = T_full (S − μ)$

5. Matched Filter (MF) #

The matched filter score is computed per pixel:

$MF = ( Sₘₙf · Xₘₙf ) / ( Sₘₙf · Sₘₙf )$

Where:

· denotes dot product.

Interpretation:

• High MF → strong projection of pixel onto target direction
• Low MF → weak or no alignment
• MF is abundance-like but not strictly constrained to [0,1].

6. Mixture Tuning – Infeasibility Metric (INF) #

To reduce false positives, MTMF evaluates whether the matched response is physically plausible under a linear mixing assumption.

First, clamp: $MF_clamped = clamp(MF, 0, 1)$

Expected mixture mean in MNF space: $μ_MT = MF_clamped · Sₘₙf$

Let $λ_bg$ be the MNF eigenvalues (background variances).

Standard deviation interpolation: $σ_MT = √λ_bg − MF_clamped ( √λ_bg − 1 )$

Variance: $λ_MT = σ_MT²$

Infeasibility: $INF = √( Σₖ ( (Xₘₙf,k − μ_MT,k)² / λ_MT,k ) )$

Properties: $INF ≥ 0$

Low $INF →$ mixture consistent with target

High $INF →$ spectrally inconsistent detection

Output Bands #

The algorithm produces two bands:

Band	Meaning
*_MF	Matched Filter response (target detection strength)
*_INF	Infeasibility metric (false-positive control)

Interpretation Guidelines #

$High MF + Low INF →$ Strong and reliable detection

$High MF + High INF →$ Spectral similarity but physically inconsistent mixture

$Low MF →$Weak or no target presence

For quantitative workflows, thresholding should consider both MF and INF.

Computational Considerations (Platform Implementation) #

To ensure scalability across large Areas of Interest (AOIs) while preserving statistical validity:

• Band cap (≤100): When an instrument provides very high band counts, bands are downselected to a maximum of 100 using a strategy that preserves spectral shape (absorption structure) and wavelength coverage.
• Bounded statistics budget: Background statistics (mean and covariance matrices) are computed using a bounded pixel budget to keep computational cost predictable across AOI sizes.
• Fused linear transforms: Noise whitening and the MNF rotation are combined into a single linear transform, reducing per-pixel matrix operations during evaluation and visualization.
• Numerical regularization: Covariance matrices are regularized with εI (ε = 10⁻⁶) to prevent instability in matrix inversion/eigen-analysis when data are near-singular or highly correlated.

These design choices allow MTMF to scale to large regions efficiently without changing the underlying mathematical formulation.

When to Use MTMF #

Mineral exploration

Detect subtle mineral signatures within mixed lithologies.

Alteration mapping

Identify weak hydrothermal alteration signals within complex backgrounds.

Hyperspectral anomaly detection

Isolate materials with known reference spectra under structured noise conditions.

High-spectral-resolution instruments (e.g., EnMAP, EMIT) yield the most robust MTMF results, though the method can run on any optical instrument with sufficient spectral sampling.