1 of 8

Atmospheric Correction:�Comparison of the Empirical Line Method with Radiative Transfer Models on Hyperspectral Data

Nikki Kozel, Electrical Engineering

Dr. Melba Crawford, Civil Engineering

1

This material is based upon work supported by the IoT4Ag Engineering Research Center funded by the National Science Foundation (NSF) under NSF Cooperative Agreement Number EEC-1941529.  Any opinions, findings and conclusions, or recommendations expressed in this material are those of the author(s), and do not necessarily reflect those of the NSF.

IoT4Ag

JULY 28, 2021

REU PRESENTATION

2 of 8

Introduction and Problem Statement

  • Image data acquired from unmanned aerial vehicles requires correction for atmospheric effects and sun angle

​

​

​

​

​

​

  • Multiple methods exist to correct for atmospheric factors, like the Empirical line method (ELM) and radiative transfer models with varying effectiveness
  • Which atmospheric correction method produces the most accurate results?

2

Source: https://www.mathworks.com/help/images/hyperspectral-data-correction.html

​

IoT4Ag

JULY 28, 2021

REU PRESENTATION

3 of 8

Overview of the Technical Approach

  • Hyperspectral images of corn and sorghum were collected with unmanned aerial vehicles
  • The data was processed with three methods: ELM, QUick Atmospheric Correction (QUAC), and Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH)
  • ELM does regression with known calibrated target values
  • QUAC determines correction based only on the image and sensor type
  • FLAASH is a radiative transfer model that takes more atmospheric inputs for correction to account for specific atmospheric conditions

​

3

Source: Dr. Melba Crawford, Purdue University

Source: https://www.spectral.com/our-resources/case-studies/success-stories-flaash/

IoT4Ag

JULY 28, 2021

REU PRESENTATION

4 of 8

Overview of the Technical Approach

  • Reflectance data from each approach is compared over soil, crops, and three calibrated targets to observe how they differ from expected values
  • The root mean square error (RMSE) for the different approaches over the calibrated targets are determined

​

4

 

The root mean square error equation

One of the images tested showing soil, crops, and the calibrated targets

IoT4Ag

JULY 28, 2021

REU PRESENTATION

5 of 8

Results

  • Images were used from two dates with different conditions. Shown here are the results for VNIR on August 6, 2020 and SWIR on June 17, 2021
  • The RMSE value over the known targets for each method were averaged

​

5

An image of the field from August 6, 2020 (right) showing the calibrated targets has the reflectance of its crops measured after correction (above). The three methods are shown: ELM (red), FLAASH (green), and QUAC (blue). There is an increase in reflectance in the green range of the visible spectrum and the near infrared range.

Atmospheric conditions for August 6, 2020 and June 17, 2021, the two dates used for analysis

IoT4Ag

JULY 28, 2021

REU PRESENTATION

6 of 8

Results

  • ELM has the lowest RMSE value, which was expected, as it used the calibrated target reference data directly
  • FLAASH performed well, although the predictions were more noisy (higher RMSE). It does not directly use the target values, as it is totally physics based

​

6

IoT4Ag

JULY 28, 2021

REU PRESENTATION

7 of 8

Results

  • While ELM has a low RMSE, the data was fitted to the calibrated target values
  • QUAC performed unusually well compared to FLAASH. Due to fewer inputs, the results typically are not as good
  • FLAASH is the most ideal method since it can account for more atmospheric variables and can make corrections without needing to take measurements of the calibrated targets

​

7

IoT4Ag

JULY 28, 2021

REU PRESENTATION

8 of 8

Summary

  • With multiple correction methods being available, knowing the accuracy of each one can help determine what equipment should be purchased for data collection
  • If the calibrated target values are unknown, QUAC and FLAASH both provide estimations of expected results, with FLAASH being more accurate
  • In the future, it would be good to see what measurement changes would make FLAASH more accurate and if more advance equipment would be needed

​

8

Source: https://www.uv.es/euroskyrad/instruments.html

IoT4Ag

JULY 28, 2021

REU PRESENTATION