Implementation of the directly-georeferenced hyperspectral point cloud

Abstract

Before pushbroom hyperspectral imaging (HSI) data can be applied in remote sensing applications, it must typically be preprocessed through radiometric correction, atmospheric compensation, geometric correction and spatial resampling procedures. After these preprocessing procedures, HSI data are conventionally given as georeferenced raster images. The raster data model compromises the spatial-spectral integrity of HSI data, leading to suboptimal results in various applications. Inamdar et al. (2021) developed a point cloud data format, the Directly-Georeferenced Hyperspectral Point Cloud (DHPC), that preserves the spatial-spectral integrity of HSI data more effectively than rasters. The DHPC is generated through a data fusion workflow that uses conventional preprocessing protocols with a modification to the digital surface model used in the geometric correction. Even with the additional elevation information, the DHPC is still stored with file sizes up to 13 times smaller than conventional rasters, making it ideal for data distribution. Our article aims to describe the DHPC data fusion workflow from Inamdar et al. (2021), providing all the required tools for its integration in pre-existing processing workflows. This includes a MATLAB script that can be readily applied to carry out the modification that must be made to the digital surface model used in the geometric correction. The MATLAB script first derives the point spread function of the HSI data and then convolves it with the digital surface model input in the geometric correction. By breaking down the MATLAB script and describing its functions, data providers can readily develop their own implementation if necessary. The derived point spread function is also useful for characterizing HSI data, quantifying the contribution of materials to the spectrum from any given pixel as a function of distance from the pixel center. Overall, our work makes the implementation of the DHPC data fusion workflow transparent and approachable for end users and data providers. Our article describes the Directly-Georeferenced Hyperspectral Point Cloud (DHPC) data fusion workflow, which can be readily implemented with existing processing protocols by modifying the input digital surface model used in the geometric correction. We provide a MATLAB function that performs the modification to the digital surface model required for the DHPC workflow. This MATLAB script derives the point spread function of the hyperspectral imager and convolves it with the digital surface model so that the elevation data are more spatially consistent with the hyperspectral imaging data as collected. We highlight the increased effectiveness of the DHPC over conventional raster end products in terms of spatial-spectral data integrity, data storage requirements, hyperspectral imaging application results and site exploration via virtual and augmented reality.

Spatial response resampling (SR2): Accounting for the spatial point spread function in hyperspectral image resampling

“…The developed MATLAB function is based on the DHPC_DSM_BLUR.m function developed in Inamdar et al. [25] . The purpose of the DHPC_DSM_BLUR.m function was to convolve a digital surface model with the PSF of a coarser resolution HSI dataset for data fusion.…”

“… [19] using the implementation from Inamdar et al. [25] . In this process, the nominal pixel size of the simulated sensor (without considering pixel summing) is first calculated based on the field of view, number of cross track pixels, integration time and speed.…”

Multi-Scale Spectral Separability of Submerged Aquatic Vegetation Species in a Freshwater Ecosystem

“…A maximum correlation coefficient threshold of 0.9 was applied to reduce the dimensionality of the DII data from 5,565 (all pairs) to 124. Following the methodology described in (Inamdar et al, 2021a;Inamdar et al, 2021b), the DII bands were geocorrected without raster resampling to generate a hyperspectral point cloud which assigns 3D spatial coordinates to each image spectrum without introducing the pixel duplications and loss that result from conventional nearest neighbour raster resampling. Next, in CloudCompare v2.12, the point cloud was subset to the study site and was rasterized to 25 cm pixels (empty cells were not interpolated over) to allow data visualization of the point cloud in raster data format without introducing pixel loss or duplication; NoData values were given to empty cells.…”

“…For the validation points recorded in situ, the buffer diameter was calculated as the sum of the manufacturer reported accuracy for NTRIP baselines >10 km (i.e., 1 m) and the average standard deviation of the points reported by the Reach RS + unit (Elmer and Kalacska, 2021). For the pixels output from the target detection, the uncertainty buffer considered the reported spatial accuracy of the CASI imagery (2.25 m) (Elmer and Kalacska, 2021) and the effective pixel resolution, the area corresponding to the full-width half -max of the CASI's point spread function (Inamdar et al, 2021a) (i.e., 1.038 m in the across-track and 0.978 m in the along-track) as determined following (Inamdar et al, 2021b). The sum of the spatial accuracy and the effective pixel resolution resulted in elliptical uncertainty boundaries where materials contributing to each pixel's recorded signal were located.…”

“…Airborne Hyperspectral Imagery Depth Invariant Index A novel DII transformation (Inamdar et al, 2021c;Lyzenga 1978, Lyzenga, 1981 based on a hyperspectral point cloud as opposed to conventional geocorrected, resampled raster imagery (Inamdar et al, 2021a;Inamdar et al, 2021b) was implemented to optimize the data quality of the CASI imagery. While Lyzenga's DII may be highly effective at distinguishing between bottom cover types, its performance depends entirely on the two input bands.…”