Hyperspectral Processing

The goal of hyperspectral data processing is to create an orthomosaic where the pixel values are apparent reflectance. This process involves two components: radiometric and geometric correction.

Radiometric Correction

Radiometric correction converts the raw data from the hyperspectral sensor to reflectance. The process comprises two steps: 1) digital number (DN) to radiance and 2) radiance to reflectance.

The figure below shows the components involved in radiometric correction, including:

• Digital number (DN): Raw data from the sensor where the pixel values have no physical meaning.

• Radiance: The amount of radiation from a given area. Radiance is the variable directly measured by the sensor. Its value is contingent on factors such as illumination, the target's position and orientation, and the path of light through the atmosphere.

• Reflectance: The proportion of radiation reflected from a surface. Reflectance is a property of materials.

Radiometric correction process.

Digital number to radiance

A gain and offset is applied to the pixel values to convert DNs to radiance. These gain and offset values are determined through a radiometric calibration of the sensor and are typically provided by the manufacturer.

Radiance to reflectance

The radiance is atmospherically corrected to derive the reflectance. There are several methods for atmospheric correction, including empirical line method (ELM), QUick Atmospheric Correction (QUAC), Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH), etc.

Empirical line method (ELM)

ELM assumes a linear relationship between radiance and reflectance. Targets with known reflectance values, typically obtained through field measurements, must be captured in the data. A linear regression (equation shown below) is used for each band to equate the radiance (data from the UAV) and reflectance (field measurements) over the targets.

$Reflectance = gain * radiance + offset$

QUick Atmospheric Correction (QUAC)

QUAC is based on the empirical finding that the average reflectance of various material spectra (excluding highly structured materials like vegetation, shallow water, and mud) remains consistent across different scenes. It is an in-scene approach and requires no additional external data. However, it is less accurate as compared to ELM and FLAASH.