Poster Session 2, Tuesday, October 4, 10:40–12:40
Leveraging mixture density networks to compensate for atmospheric effects over inland and coastal waters
The primary challenge for the remote sensing of aquatic ecosystems via modern remote sensing systems revolves around formulating and implementing a robust atmospheric correction (AC) model to predict remote sensing reflectance (R rs ). This presentation describes the concept of a machine-learning model implemented for estimating R rs from Landsat-8’s Operational Land Imager’s (OLI) observations over inland and coastal waters. First, a well-validated, coupled ocean-atmosphere radiative transfer (RT) model was adapted to integrate into situ R rs as the surface boundary condition. OLI’s top-of-atmosphere reflectance spectra (rhot) were simulated for a plausible range of imaging geometries and aerosol contents, as well as for a broad range of in situ R rs spectra. Second, a class of neural networks, i.e., mixture density networks that model the conditional probability of input-output variables, was designed and tested to transform OLI’s simulated rhot to R rs . The model was then applied to OLI imagery for an extensive matchup assessment. Compared to other existing AC processors, our matchup assessments (N > 500) showed performance improvements from 20 to more than 100%, depending on the spectral band, statistical metrics, and benchmark AC processors. Median uncertainties for the performance of our model were 16%, 15%, 19%, and 30% for OLI’s 443, 482, 560, and 655 nm spectral bands, respectively. Our model notably exhibited superior performance in reducing retrieval noise, as evidenced by the root-mean-square statistics relative to benchmark processors. These preliminary results suggest that our methodology is a promising approach to the inverse problem of AC.
Akash Ashapure, SSAI, [email protected]
Brandon Smith, SSAI, [email protected]
Pengwang Zhai, University of Maryland Baltimore County, [email protected]