Atmospheric correction intercomparison of hyperspectral and multispectral imagery over agricultural study sites

In this research effort we assess the performance of atmospheric correction-based surface reflectance (SR) retrievals from two satellite image sources, one with very high spatial resolution (VHR) (<5-m) and the other high spectral resolution (~10-nm). The VHR images are from MAXARs WorldView-3 (WV3) satellite and the high spectral resolution images are from Agenzia Spaziale Italianas (ASI) PRecursore IperSpettrale della Missione Applicativa (PRISMA) satellite. We use various atmospheric correction (AC) tools to provide intercomparisons of both AC tools and image source SR estimates. The AC tools we evaluated include Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) within ENVI version 4.7, MODerate resolution atmospheric TRANsmission (MODTRAN) versions 5.3.3 and 6.0, and ASIs Level-2D correction for PRISMA imagery. Prior to correcting WV3 and PRISMA imagery to SR, we performed manual geometric corrections of imagery as both image sources were found to lack consistent georegistration.

We performed comparisons at two study sites in Maryland, USA, including the United States Department of Agriculture Beltsville Agricultural Research Center (BARC) and an agricultural study site on Marylands Eastern Shore region. For the BARC site, we used WV3 imagery acquired on 2022-04-02 and PRISMA imagery acquired on 2022-04-28, focusing on evaluation of AC tool SR retrieval performance for each image source separately due to large time differences in image acquisitions where SR values are likely impacted by changing field conditions. For the Eastern Shore site, WV3 imagery was acquired on 2022-05-18 and 2022-05-30, and PRISMA imagery was acquired on 2022-05-21, allowing for quantitative evaluation of both AC tool performance and intercomparison between WV3 and PRISMA imagery. Having WV3 imagery acquired before and after PRISMA imagery allows for interpretation of major changes in field conditions and thus, identification of fields to exclude from intercomparisons. For intercomparison assessments, we computed relative percent difference (RPD) between the AC tool SR retrievals. For image source comparisons, 4-m WV3 pixels were resampled to 30-m PRISMA pixels after which 30-m WV3 bands and PRISMA spectra were compared to one another visually for both study sites. To provide rigorous SR retrieval intercomparisons between image sources, PRISMA spectra were resampled to WV3-equivalent bands for RPD computation for the Eastern Shore site.

In addition to the SR retrieval intercomparisons between the AC tools, we carry out a quasi-validation where we retrieve fractional crop residue cover (fR) from the satellite image sources by calculating established spectral indices (SIs) and calibrating SIs with ground-measured fR acquired within several days of satellite overpasses. These SIs include the Cellulose Absorption Index (CAI) (Nagler et al. 2000), Shortwave Infrared Normalized Difference Residue Index (SINDRI) (Serbin et al. 2009), Lignin-Cellulose Absorption Index (LCAI) (Daughtry et al. 2005), and Lignin-Cellulose Peak Center Difference Index (LCPCDI) (Hively et al. 2021) 1-4. The most accurate crop residue SIs are generally based on shortwave infrared (SWIR) reflectance bands ranging from 2000 nm to 2400 nm that measure dry vegetation lignocellulose absorption features at 2100 and 2300 nm 1-5. For instance, the CAI identifies a 2100 nm cellulose absorption feature with a central band positioned on this feature, and two spectrally adjacent bands at 2040 and 2210 nm, while the LCAI identifies the 2300 nm lignin absorption feature compared to bands at 2165 and 2210 nm. Particular focus on intercomparisons for the SWIR region is critical as atmospheric water, carbon dioxide, and methane impact accurate SR retrieval as shown in Figure 1.a. Our final analysis concludes with the selection of the top-performing AC approach between the WV3 and PRISMA imagery (as indicated by low SR RPD) and then compares PRIMSA and 30-m WV3 imagery with original 4-m WV3 imagery to assess the degree to which spatial resolution impacts the retrieval of fR. Figure 1 provides a comparative example of WV3 and PRISMA imagery used to compute SINDRI which is then calibrated to fR using second order polynomial equations from Hively et al. (2018) 6. Figure 1 fR calibrations will be updated with newly acquired ground survey data from May 2022 to further improve the accuracy of image source and AC tool intercomparisons.