Water column compensation workflow for hyperspectral imaging data

Abstract

Our article describes a data processing workflow for hyperspectral imaging data to compensate for the water column in shallow, clear to moderate optical water types. We provide a MATLAB script that can be readily used to implement the described workflow. We break down each code segment of this script so that it is more approachable for use and modification by end users and data providers. The workflow initially implements the method for water column compensation described in Lyzenga (1978) and Lyzenga (1981), generating depth invariant indices from spectral band pairs. Given the high dimensionality of hyperspectral imaging data, an overwhelming number of depth invariant indices are generated in the workflow. As such, a correlation based feature selection methodology is applied to remove redundant depth invariant indices. In a post-processing step, a principal component transformation is applied, extracting features that account for a substantial amount of the variance from the non-redundant depth invariant indices while reducing dimensionality. To fully showcase the developed methodology and its potential for extracting bottom type information, we provide an example output of the water column compensation workflow using hyperspectral imaging data collected over the coast of Philpott's Island in Long Sault Parkway provincial park, Ontario, Canada.•Workflow calculates depth invariant indices for hyperspectral imaging data to compensate for the water column in shallow, clear to moderate optical water types.•The applied principal component transformation generates features that account for a substantial amount of the variance from the depth invariant indices while reducing dimensionality.•The output (both depth invariant index image and principal component image) allows for the analysis of bottom type in shallow, clear to moderate optical water types.

Keywords: Depth invariant index (DII); Hyperspectral imaging; Principal component analysis (PCA); Water column compensation.

Graphical abstract

Fig. 1

Reflectance spectra from Hyperspectral Imaging data observed over a flooded cement road at the Long Sault Parkway in Ontario, Canada at various water depths. As the water depth increases, the observed reflectance is attenuated.

Fig. 2

Flow chart of the water column compensation workflow for hyperspectral imaging (HSI) data. The workflow calculates depth invariant indices (DII) for each band pair from the original HSI data. Due to the high dimensionality of HSI data, the number of calculated DIIs are reduced using a correlation based feature selection algorithm. The data dimensionality is additionally reduced using a principal component transformation (PCA). The DII and PCA end products are highlighted with a red boarder in the workflow.

Fig. 3

Photograph of Philpott's Island in Long Sault Parkway provincial park. The photograph was captured with the X5s camera aboard a DJI Inspire 2 remotely piloted aerial system.

Fig. 4

(A) The CASI hyperspectral imaging data collected over the Long Sault Parkway (red = 639.6 nm, green = 548.7 nm, blue = 472.1 nm, linearly stretched between 0 and 13%). Before being input into the water column compensation workflow, the land pixels needed to be masked from the imagery. (B) The CASI hyperspectral imaging data (red = 639.6 nm, green = 548.7 nm, blue = 472.1 nm, linearly stretched between 0 and 3%) after masking out land pixels. The regions of interest (ROIs) required by the workflow are shown in the figure in green (deep water pixels) and red (pixels along a transect of varying water column depths over a constant substrate).

Fig. 5

Example output of the water column compensation workflow. (A) Input CASI hyperspectral imagery over Long Sault Parkway (red = 639.6 nm, green = 548.7 nm, blue = 472.1 nm, linearly stretched between 0 and 3%). (B) DII data product (red = DII 75 (682.7 nm & 701.8 nm), green = DII 22 (553.5 nm & 563.1 nm), blue = DII 7 (424.3 nm & 438.7 nm), linearly stretched from minimum to maximum value on extent). (C) PCA transformed DII data product (red = PC 1, green = PC 2, blue = PC 3, linearly stretched from minimum to maximum value on extent).

Fig. 6

Example application of the water column compensation workflow and resultant class spectra and plots demonstrating the improved performance of the DII and PCA as compared to original hyperspectral imagery. (A) High resolution RGB photograph of the study area (22 m by 22 m region) captured with an X5s camera aboard a DJI Inspire 2 remotely piloted aerial system. (B) Manual delineation of bottom cover classes defined from the orthomosaic and in situ field knowledge of the site. (C) Input CASI hyperspectral image over Long Sault Parkway (red = 644.4 nm, green = 548.7 nm, blue = 472.1 nm, linearly stretched from minimum to maximum value on extent). (D) Plot of mean class spectra extracted from the airborne hyperspectral image. (E) DII data product (red = DII 75 (682.7 nm & 701.8 nm), green = DII 22 (553.5 nm & 563.1 nm), blue = DII 7 (424.3 nm & 438.7 nm), linearly stretched from minimum to maximum value on extent). (F) Plots of the DII values of each class in the study area for DII bands 28 (572.63 nm + 577.42 nm), 30 (596.56 nm + 601.35 nm) and 31 (606.13 nm + 610.92 nm). The DII bands were selected as they had the greatest separability between the two vegetation classes (absolute difference between the mean DII for each class, normalized by the product of the standard deviation of the DII for each class). (G) PCA transformed DII data product (red = PC 1, green = PC 2, blue = PC 3, linearly stretched from minimum to maximum value on extent). (H) Plots of the PCA values of each class in the study area for the first three principal components. The legend in subplot B is applicable to subplots D, F and H.