Optimized spectral indices for camouflage detection in multispectral imagery

Unmanned aerial vehicles, equipped with multispectral imaging systems, as well as publicly available datasets, have fueled various research in remote sensing applications, including precision agriculture, land cover mapping or even camouflage detection. Many of these applications make use of spectral indices, such as normalized difference or ratio indices, created by merging multiple raw bands. These indices typically provide a direct indication of certain physical surface properties, like plant health, nitrogen content or leaf area index, and are determined by studying spectral reflectance properties or using optimization techniques. Given the heavy use and utility of such indices, this work introduces a novel generalization of the normalized difference, ratio and difference indices, the linear ratio index (LRI), a ratio of two linear functions of all available bands. In addition, an optimization approach for the LRI is presented, which incorporates complexity reduction strategies that enable optimization using only a subset of all available bands and reduced parameter precision, thereby ensuring parameter readability. The LRI and its optimization are thoroughly investigated in the context of camouflage detection in tactical reconnaissance scenarios by optimizing a six-band and a two-band LRI using the eXtended Multispectral Dataset for Camouflage Detection (MUDCAD-X). For comparison with traditional spectral index optimization approaches, the resulting linear ratio indices (LRIs) are evaluated against all raw bands and an optimized normalized difference index and an optimized ratio index. The evaluation shows that the optimized LRIs provide the best overall results in terms of visibility and detectability of camouflaged targets. This could indicate a general superiority of the LRI over established indices optimized by testing band permutations, making the LRI a promising candidate for further investigation in other remote sensing applications where it could also outperform traditional index optimization approaches. Therefore, the software code for optimizing the LRI has been made publicly available for further exploitation, requiring only an adapted optimization criterion to support any other use case. Furthermore, as with most spectral indices, the LRIs obtained in this study have negligible computational overhead and, under the right conditions, can be directly integrated into any existing camouflage detection system.