Abstract
In land-based hyperspectral imaging (HSI), the spectral signatures of ground objects are significantly influenced by varying environmental conditions, such as solar position, weather, and shadows. This leads to the "same object, different spectrum" phenomenon, which degrades the performance of traditional target detection (TD) algorithms. This paper introduces the Weighted Spectral Correlation Angle (WSCA) method, designed to overcome these spectral uncertainties. By partitioning the spectrum into stable "common bands" and volatile "feature bands" and applying a weighting strategy, WSCA enhances detection reliability. Experimental results demonstrate that WSCA outperforms traditional methods like Spectral Angle Mapper (SAM) and Constrained Energy Minimization (CEM), maintaining high detection accuracy across diverse temporal and spatial conditions.
Introduction: The Challenge of Spectral Uncertainty
While hyperspectral signatures are often treated as unique "barcodes" in controlled laboratory settings, field-based imaging introduces significant noise. Factors such as the solar zenith and azimuth angles, atmospheric conditions, and the spatial orientation of the target cause the observed reflectance to fluctuate. Traditional TD methods typically rely on rigid similarity measures that fail when the target's spectral curve is distorted. This often results in false positives or missed detections, particularly in complex environments where targets like camouflage are designed to blend into the background.
The Weighted Spectral Correlation Angle (WSCA) Method
To address these limitations, the authors developed the WSCA method, building upon the foundations of the Spectral Correlation Angle (SCA). SCA is inherently robust against multiplicative brightness changes and additive offsets, making it a strong baseline for land-based imaging. However, standard SCA treats all spectral bands equally, which is suboptimal when certain wavelengths are more prone to environmental interference than others.
The WSCA approach introduces a strategic weighting mechanism:
- Band Division: The full spectrum is analyzed to identify "feature bands" (high variance under different conditions) and "common bands" (relatively stable).
- Weighting ($k$): A weight coefficient, $k$, is applied to the stable common bands.
- Similarity Enhancement: By mathematically prioritizing the stable portions of the spectrum, the algorithm effectively filters out the noise present in the uncertain bands, leading to a more consistent similarity score between the reference and the observed spectra.
Experimental Setup
The effectiveness of WSCA was evaluated using a land-based HSI-300 spectrometer capturing 89 bands (449–801 nm). The test scene involved camouflaged clothing placed against a background of sycamore leaves, soil, and weeds.
Two primary experiments were conducted:
- Temporal/Environmental Variation: Data was collected at 30-minute intervals over several hours to capture the impact of changing solar angles.
- Spatial Variation: The target was analyzed across different subregions to test robustness against uneven spatial distribution and self-shadowing.
Results and Discussion
The WSCA method was benchmarked against five standard algorithms: SAC, NED, NCC, CEM, and MI. Performance was quantified using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values.
- Environmental Stability: In the temporal experiment, WSCA achieved an average AUC of 0.9978, significantly higher than the Spectral Angle Cosine (0.9763) and Mutual Information (0.9709) methods.
- Spatial Robustness: WSCA maintained a near-perfect AUC of 0.9986 in the spatial variation tests, proving its ability to handle targets that are not spectrally uniform.
- Qualitative Performance: While methods like CEM provide high contrast, they lack the stability of WSCA across varying imaging conditions, making WSCA a more reliable choice for persistent monitoring.
Conclusion
The Weighted Spectral Correlation Angle method provides a robust solution for target detection in the presence of spectral uncertainty. By focusing on the most stable parts of the hyperspectral signature, it bridges the gap between theoretical spectral libraries and practical field applications. Future work will focus on automating the band selection process and optimizing the weighting coefficient $k$ for different material classes and environmental contexts.
Summary updated