An Efficient Machine Learning Approach for Atmospheric Correction

Prankur Rusia, S Manthira Moorthi

Abstract:

The earth observation using remote sensing images is an inquisitive way to explore and evaluate
the geo-resources of any specific area on the globe. In this regard, Indian Resourcesat-2A (R2A)
remote sensing satellite plays an important role in monitoring the critical resources of our planet
using its unique three tier imaging mechanism. Optical sensors on-board R2A have good spatial
temporal resolution for diverse space borne applications. Most of these applications requires Surface
Reflectance (SR) data product by removing the effects of intermittent atmospheric scattering and
absorption. Radiative Transfer Models (RTM) are used to perform atmospheric correction which are
computationally intensive, thus a Look-Up-Table (LUT) is utilized to interpolate intermediate values
as a trade-off between accuracy and speed. However, the process of interpolation too becomes very
compute intensive when a large enough LUT is used. The paper provides an approach to remove this
trade-off by using multi-layered deep network to model interpolation as a regression problem. The
proposed method generates highly accurate Deep SR product with a significant reduction in turnaround time. The experimental result shows that a speedup of 5x is achieved with the developed
framework as compared to conventional interpolation-based approach for generation of R2A LISS-3
Deep SR scene data product. The Deep SR product is compared with pure 6SV generated product
and R2 value found to be 0.97 (Green), 0.97 (Red), 0.98 (NIR) and 0.98 (SWIR) respectively. To
check the efficacy of the framework, the LISS-3 Deep SR product is also compared with closest
acquisition Landsat-8 SR product and ground truth values obtained through vicarious calibration.
The maximum relative deviation error found to be 1.34%, 1.82%, 3.25% and 2.16% for Green, Red,
NIR and SWIR channels respectively.