Abstract

High-resolution water quality maps derived from imaging spectroscopy provide valuable insights for environmental monitoring and management, but the processing of all pixels of large datasets is extremely computationally intensive and limits the speed of map production. We demonstrate a superpixel approach to accelerating water quality parameter inversion on such data to considerably reduce time and resource needs. Neighboring pixels were clustered into spectrally similar superpixels, and bio-optical inversions were performed at the superpixel level before a nearest-neighbor interpolation of the results back to pixel resolution. We tested the approach on five example airborne imaging spectroscopy datasets from Hawaiian coastal waters, comparing outputs to pixel-by-pixel inversions for three water quality parameters: suspended particulate matter, chlorophyll-a, and colored dissolved organic matter. We found significant reduction in computational time, ranging from 38 to 2625 times faster processing for superpixel sizes of 50 to 5000 pixels (200 to 20,000 m2). Using 1000 paired output values from each example image, we found minimal reduction in accuracy (as decrease in R2 or increase in RMSE) of the model results when the superpixel size was less than 750 2 m × 2 m resolution pixels. Such results mean that this methodology could reduce the time needed to produce regional- or global-scale maps and thereby allow environmental managers and other stakeholders to more rapidly understand and respond to changing water quality conditions.