An investigation into reach scale estimates of sub-pixel fluvial grain size from hyperspatial imagery.

Abstract

Grain size data for gravel bed rivers is important in a wide variety of contexts; providing crucial information to guide the development of flood defences, and maintaining navi-gability, biodiversity and ecological integrity within large gravel bed rivers. Advances in remote sensing technologies have seen an increase in the acquisition of hyperspatial imagery (imagery with a spatial resolution of < 10 cm), and advances in computational power have complemented this data acquisition allowing for the application of complex image processing techniques. An improved methodology is presented for extracting reach scale grain size information. Of particular note is the ability to generate estimates of sub-pixel surface sand content, as well as sub-pixel grain size distributions. The methodology was applied to Queens Bar, NBar, Calamity Bar and Harrison Bar within the gravel reach of the Fraser River (British Columbia, Canada).

Hyperspatial imagery was acquired at 3 cm resolution, along with independent surface grain size information. Surface sand estimates were calculated through a first order standard deviation textural layer; calibrations revealed an inequality based relationship be-tween texture and sand content, allowing for the production of binary maps of surface sand content with an approximate accuracy of 70%. Calibrations were calculated for 7 grain size percentiles for the gravel fraction of the grain size distribution ( > 2 mm); D5, D16, D35, D50, D65, D84 and D95 were achieved, following a wide ranging parameter investigation. A combination of first order standard deviation along with several second order Grey Level Co-occurrence Matrix textural parameters (entropy, contrast and cor-relation) calibrated to grain size using multiple linear regression. The best performing calibrations were found for smaller and intermediate percentiles; cross validated mean square error (%) at 0.61, 3.55, 9.58 and 16.25 for D5, D16, D35, and D50 respectively. Calibrations began to break down for the largest percentiles; cross validated mean square error (%) at 26.43 and 44.99 for D84 and D95. The breakdown of calibrations for larger percentiles is attributed to the ‘pixel averaging eect’; for smaller percentiles a larger population of grains were averaged into one pixel, thus variance across multiple pixels is low, whereas for the larger percentiles the grain size approaches the spatial resolution of the pixels, therefore a smaller population of grains makes up one pixel and introduces in-creased variance across multiple pixels. Overall, this new methodology presents a means for extracting sub-pixel grain size information from hyperspatial imagery, with higher ac-curacies for the smaller percentiles than previously published. This allows for the rapid acquisition of a large amount of grain size information without the need for time intensive field techniques.