SPIE Membership Get updates from SPIE Newsroom
  • Newsroom Home
  • Astronomy
  • Biomedical Optics & Medical Imaging
  • Defense & Security
  • Electronic Imaging & Signal Processing
  • Illumination & Displays
  • Lasers & Sources
  • Micro/Nano Lithography
  • Nanotechnology
  • Optical Design & Engineering
  • Optoelectronics & Communications
  • Remote Sensing
  • Sensing & Measurement
  • Solar & Alternative Energy
  • Sign up for Newsroom E-Alerts
  • Information for:
SPIE Photonics West 2019 | Register Today

SPIE Defense + Commercial Sensing 2019 | Call for Papers

2019 SPIE Optics + Photonics | Call for Papers



Print PageEmail PageView PDF

Remote Sensing

Neural nets enhance atmospheric data from infrared images

Atmospheric information from satellite images can successfully be estimated by neural networks, despite intruding effects.
21 March 2008, SPIE Newsroom. DOI: 10.1117/2.1200802.1048

In the infrared domain, terrestrial surfaces give off radiant energy according to their spectral emissivity and surface temperature. Remote infrared hyperspectral imagery provides new opportunities to exploit these phenomena both to discriminate surface material and to obtain information about a surface's thermal state. The infrared is of particular interest for night observation and for geological studies, as rock emissivities exhibit specific spectral features.

At-sensor radiance is polluted by the atmosphere as it absorbs and emits radiation mainly in absorption bands but also in atmospheric windows (areas that offer optimal observing conditions, such as clear skies). Assuming that the surface is Lambertian—that is, independent of the incidence of radiation and the direction of observation—the spectral radiance reaching an airborne sensor can be written as:

where τ is the atmospheric transmittance, Eatm the downward irradiance, Latm the upward atmospheric radiance, B the Planck function, and Tsurf the surface temperature. In the 8–12μm range and under clear sky conditions, atmospheric variables depend mostly on water vapor concentration and air temperature. To use this formula to best effect, the process by which each of the variables is obtained, from both the infrared spectral data and from the data output of added neural networks (NNs), must be understood.

First, spectral emissivity involves the retrieval of information on atmospheric compensation, surface temperature, and emissivity separation. To ensure the quality of that information, we have developed a new atmospheric compensation method suited to a future airborne hyperspectral sensor.1 The sensor operates within the 3–5.5 and 8–12μm spectral band. The method combines NNs to retrieve atmospheric temperature and water vapor content. An additional spectral smoothness technique is used for temperature emissivity separation. The efficiency of this smoothness technique has already been proven in the long-wave infrared (LWIR) domain.2

The NN was programmed and instructed to retrieve the mean temperature of the atmospheric layer under the aircraft, and a second NN was created to retrieve the columnar water vapor (CWV) from the same location. This second NN was created based on the experience of other researchers who had used NNs for atmospheric sounding from satellite instruments, such as the TIROS Operational Vertical Sounder, the Atmospheric InfraRed Sounder, and IASI (Interféromètre Atmosphérique de Sondage Infrarouge), and from airborne instruments at 8km altitude.3–6

The inputs of the first NN, the mean temperatures, were radiances computed with MODTRAN4 in the CO2 absorption band between 4 and 4.5μm with a 16cm−1 spectral resolution. The inputs of the second NN, the CWV, were radiances between 4.8 and 5.5μm. This choice of wavelength represents the edge of the water vapor absorption band and the mean atmospheric temperature obtained with the first NN.

The quality of an NN depends on the learning step and especially on how representative the learning data set is. Simulations were carried out using Aster emissivities and the TIGR (Thermodynamic Initial Guess Retrieval) atmospheric database,3 which is an exhaustive sampling of all possible atmospheres. Results obtained for the TIGR atmospheric class ‘midlatitude2’ (mainly summer) and solar zenith angle in the range of 20–40° are shown on Figure 1. The retrievals obtained with an independent validation data set were compared with the true values of the mean temperature and CWV.

Figure 1. Mean atmospheric temperature and water vapor content retrieved with NNs vs. their true values. Dashed lines on the temperature plot (left) stand for true value ± 2K. On the water vapor plot (right), the dashed lines correspond to the true value ± 10%.

For an ideal instrument, with a high signal-to-noise ratio (SNR) and perfect spectral and radiometric calibration, the method introduced has been able to retrieve the mean atmospheric temperature with an accuracy of about 0.3K. CWV was also retrieved with an accuracy compatible with present requirements (i.e., emissivity and surface temperature estimation), except for a very few cases.

The next step consisted in retrieving the spectral emissivity and the surface temperature. Atmospheric terms in Equation 1 were computed with MODTRAN4, using mean atmospheric profiles scaled to temperature and WVC obtained with NNs. Both Tsurf and ε(λ) were estimated using the spectral smoothness algorithm2 in the 8–12μm band. Table 1 gives the results of the entire process and shows sensitivity to random noise and to a global bias of 5% radiance.

Table 1. Atmospheric compensation and surface temperature, along with emissivity separation results. SNR: Signal-to-noise ratio. Stdv: Standard deviation.

Certain factors must be taken into account when using the method. Biased radiance spectra induce atmospheric temperature and surface temperature biases. In cases where bias is +5% (of the same sign) in both 3–5.5μm and 8–12μm ranges, bias on atmospheric temperature tends to temper surface temperature bias. In this case, emissivity spectra should show degraded but still acceptable accuracy, as with a 200 SNR. However, full results1 presuppose that retrieval quality can dramatically decrease depending on the degree of radiance spectra bias.

The advantage of this atmospheric compensation approach is that it is very flexible. It can easily be adapted to various spectral imagers (different spectral resolution, flight altitude, and so on), by learning new networks with new sets of inputs and outputs. For instance, testing for images has already been carried out at 8km altitude,6 and will be soon adapted to the AIRIS (Airborne InfraRed Imaging Spectroscopy) instrument with its higher spectral resolution. Specialization according to terrain types, latitude range, and so forth can also be applied to increase performance.

Véronique Achard, Sébastian Lesage, Laurent Poutier 
Theoretical and Applied Optics Department (DOTA)
ONERA (French Aerospace Lab)
Toulouse, France

Véronique Achard is a research engineer. Her research activities currently concern hyperspectral imagery, including direct modeling, atmospheric correction, and image processing.

Sébastian Lesage is completing his PhD. He is working on atmospheric correction techniques for infrared hyperspectral imagery.

Laurent Poutier is a research engineer. He is involved in developing direct radiative transfer codes and inverse techniques for retrieving optical properties from ground, airborne, and spaceborne spectroradiometric measurements.