The Spectral Mixture Residual: A Source of Low‐Variance Information to Enhance the Explainability and Accuracy of Surface Biology and Geology Retrievals. Issue 2 (18th February 2022)

• Record Type: Journal Article
• Authors: Sousa, D.; Brodrick, P.; Cawse‐Nicholson, K.; Fisher, J. B.; Pavlick, R.; Small, C.; Thompson, D. R.
• Is Part Of: Journal of geophysical research. Volume 127:Issue 2(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-02-18
• Subjects: imaging spectroscopy -- spectral mixture -- dimensionality -- explainability -- spectral feature space -- continuum removal; Geobiology -- Periodicals; Biogeochemistry -- Periodicals; Biotic communities -- Periodicals; Geophysics -- Periodicals; 577.14
• DOI: 10.1029/2021JG006672
• Languages: English
• ISSNs: 2169-8953

Abstract

Abstract: Spectrally mixed pixels are the rule, not the exception, in decameter terrestrial imaging. By definition, the reflectance spectrum of a mixed pixel is a function of more than one generative process. Physically based surface biology and geology retrievals must therefore isolate the component of interest from a myriad of unrelated processes, each occurring with unknown presence and abundance across the hundreds of square meters sampled by each pixel. Foliar traits, for example, must be isolated from canopy structure and substrate composition. In many cases, these unrelated processes can dominate overall variance of spatially integrated reflectance. We propose a new approach to isolate low‐variance spectral signatures in mixed pixels. The aggregate effects of (high‐variance) spatial mixing processes within each pixel are modeled by treating the observed reflectance as a linear mixture of a small set of generic endmember spectra. Spatial mixing effects are removed by computing the (low‐variance) difference between the modeled and observed spectra, named the Mixture Residual (MR). The MR, a residual reflectance spectrum that is presumed to carry the subtler and variable signals of interest, is then leveraged as a source of signal. We illustrate the approach using three independent collections of reflectance spectra: synthetic composites computed from field measurements, NEON AOP airborne image compilations, and DESIS satellite data. The MR is found to discriminate between land cover versus plant trait signals, and to accentuate subtle absorption features. Mean band‐to‐band correlations within the visible, NIR, and SWIR wavebands decrease from 0.97, 0.94, and 0.97 to 0.95, 0.04, and 0.31 (respectively). The number of dimensions required to explain 99% of image variance increases from 4 to 13. We focus on vegetation as an illustrative example, but note that the concept can be extended to other classes of spectra and used as an input to other algorithms. Plain Language Summary: Imaging spectroscopy measures a rich superposition of signals generated by a wide range of biogeophysical processes and materials. In many cases, subpixel mixing—variations in the abundance of broad land cover types like soil, foliage, and shadow—can dominate these spectra at the expense of subtler signals like leaf chemistry and soil composition. This spatial mixing problem can seriously confound surface biology and geology retrieval generalizability and accuracy. We propose a way to isolate the subtle signatures of interest from strong land cover mixing signals through a new approach called the mixture residual (MR). This approach models and removes the land cover mixing signal in a highly general, computationally efficient, and invertible way. We show the effectiveness of the method using synthetic imagery computed from mixtures of field spectra, a broad compilation of ecologically diverse airborne data, and satellite imagery. This approach can complement established algorithms by adding a degree of explicability, as well as improving generalizability and classification and regression results in some cases. Key Points: Mixture residual spectrum can contain informative low‐variance signals Average band‐to‐band correlations drop from 0.94 and 0.97 to 0.04 and 0.31 for near infrared and shortwave infrared, respectively Linear approach offers feature accentuation, interpretability, generalizability, and computational efficiency