Spectral unmixing of coral reef benthos under ideal conditions

Hyperspectral remote sensing has shown promise for detailed discrimination of coral reef substratum types, but, by necessity, it samples at pixel scales larger than reef substratum patch sizes. Spectral unmixing techniques have been successful in resolving subpixel areal cover in terrestrial environments. However, the application of spectral unmixing on coral reefs is fundamentally challenging, due not only to the water column, but also to the potentially large number of spectrally similar and ecologically significant end-member (substratum) classes involved. A controlled ex-situ experiment was conducted using field-spectrometer data to assess the accuracy of spectral unmixing techniques to estimate the areal cover of small-scale (<0.25 m2) assemblages of reef substrata (e.g., changes in cover between massive corals, branching corals, bleached corals, macroalgae, and coralline red algae). Mixture compositions were obtained precisely by analysis of digital images collected by a camera calibrated to the field of view of the spectrometer. Linear unmixing techniques were applied to derive estimates of substratum proportions using the full spectral resolution data and various transformations of it, including derivatives and down sampling (merging adjacent wavelengths into broader spectral bands). Comparison of actual and estimated substratum proportions indicate that spectral unmixing may be a practical approach for estimating subpixel-scale cover of coral reef substrata. In the most accurate treatment, coefficients of determination across all mixture sets were high for most end-member classes (37 of 52 cases with r2 >0.64, i.e. r >0.8). The most successful analyses were based on derivatives of down-sampled data, implying that spectral unmixing benefits more from spectral smoothing and judicious choice of band locations than from high spectral resolution per se. Although these results show that changes in coral and algal cover can be determined by unmixing their spectra, the method is not yet an operational remote sensing tool. Primary empirical research is needed before taking the next step, which is to incorporate a water column, of variable depth, above the sea bed.