Influence of atmospheric modeling on spectral target detection through forward modeling approach in multi-platform remote sensing data

Abstract

Identifying objects or pixels of interest that are few in numbers and sparsely populated in imagery is referred to as target detection. Traditionally, the inverse modeling (IM) approach, usually a slow and computationally intensive process, is used for detecting targets using surface reflectance spectra. For the emerging online methods in remote sensing, modeling the at-sensor radiance of target material, i.e., a forward modeling (FM) approach, can be used. Compared to the IM approach, FM is better suited to online methods due to its potential for adaptation to regional atmospheric modeling. Spectral knowledge transfer of a target from a known to an unknown atmospheric condition is the primary outcome of an efficient target detection framework. However, such an endeavor requires an exhaustive assessment of the target detection process under different atmospheric models and associated uncertainties. The objective of this work is to assess the quantitative impact of atmospheric parameters on the detectability of engineered targets. Specifically, the impact of critical atmospheric parameters such as aerosol optical thickness (AOT), standard atmospheric profiles, and aerosol models are considered. For this effect, we designed a multi-platform image acquisition setup that acquired targets concurrently using a ground-based terrestrial hyperspectral imager (THI), an airborne hyperspectral imager (AVIRIS-NG), and a space-borne multispectral imager (Sentinel-2). We used a point-based spectroradiometer and pixel-based THI to collect the in-situ reference target reflectance spectra and generated a radiance spectral library by simulating TOA radiance spectra using the Second Simulation of the Satellite Signal in the Solar Spectrum (6S) radiative transfer model. We have considered two cases of target radiance simulations, i.e., (i) corresponding to a grid of different AOT values for a predefined atmospheric and aerosol profile, and (ii) corresponding to varying combinations of atmospheric and aerosol profiles at a given AOT. The detection has been carried out using multiple target detection algorithms. Results indicate that the spectral knowledge-based targets can be detected in remote sensing data under different atmospheric model scenarios using the FM approach. A detection rate of about 75% and 50% have been consistently obtained for remote sensing data from airborne and space-borne platforms with a false alarm (FA) rate of 10-2 to 10-3 respectively. Change in the AOT across atmospheric models has resulted in decision-changing implications in the target detection modeling. The selection of the wrong atmospheric profile can potentially aggravate the number of FAs produced by a particular detection algorithm.