LECTURE 6

Summary

The lecture continues the discussion of hyperspectral imagery and introduces synthetic aperture radar (SAR), focusing on how each type of data supports Earth-observation tasks.

Hyperspectral Imagery

Hyperspectral images contain many closely spaced spectral bands. Their detailed spectral signatures can help identify materials, but similar materials may have overlapping signatures, so classification depends on how separable their spectra are. Using spatial context alongside spectral information can help distinguish classes; the lecturer says this will be explored in later course material.

Important hyperspectral analysis tasks include:

• Anomaly detection: Find spectrally unusual pixels or objects without a known target signature. An anomaly is rare compared with the surrounding background and may be very small; some research datasets contain only a few anomalous pixels.
• Target detection: Search for a target using a known spectral signature. This differs from anomaly detection, which does not assume the target’s signature is known.
• Unmixing: Decompose a pixel containing a mixture of materials into its component spectra, called endmembers, and estimate each material’s proportion, called its abundance.
• Enhancement: Improve data quality or usability through methods such as denoising, super-resolution, and inpainting.

Anomaly detection is used in contexts such as defense and surveillance, oil-spill detection, mineral exploration, and carbon monitoring. It can be difficult to avoid both missed detections and false alarms. A practical approach may be to detect broadly first, then reduce false alarms through post-processing—for example, using a land-cover map to exclude pixels in areas where the target is unlikely.

Dimensionality Reduction

Neighboring hyperspectral bands are often highly correlated. Keeping every band can increase computational cost, while reducing the data can make processing and visualization easier. Whether dimensionality reduction improves classification depends on the application: removing bands indiscriminately may also discard useful information.

Two main approaches are:

• Band or feature selection: Choose a subset of the original bands, for example by ranking bands and retaining the most informative ones. This preserves their original physical meaning and can remove redundant or noisy bands.
• Feature extraction or transformation: Combine information from multiple bands to create a compact set of new features. This can improve separability and reduce dimensionality, but may make the results less physically interpretable. Indices such as NDVI and NDBI are examples of transformations whose meanings remain relatively clear.

Hyperspectral Denoising

Hyperspectral noise can vary across bands. Denoising aims to remove unwanted signal perturbations while retaining spectral fidelity. A simple spatial filter, such as a mean filter, may blur spatial detail and does not adequately model spectral relationships. The lecture therefore emphasizes the value of accounting for spectral context, including in learning-based methods.

Synthetic Aperture Radar (SAR)

SAR is an active remote-sensing system: it transmits microwave pulses and measures the energy returned from Earth’s surface. Unlike optical imagery, it does not rely on sunlight, can collect data at night, and can operate through cloud cover. SAR pixels represent backscatter, rather than reflected sunlight.

The term synthetic aperture refers to using a sensor’s movement and measurements from a shorter antenna to simulate the effect of a longer antenna, helping achieve finer resolution.

SAR appearance depends strongly on surface properties and viewing geometry:

• Smooth surfaces, such as calm water, often reflect energy away from the sensor and appear dark.
• Rough surfaces scatter energy in more directions and often appear brighter.
• Vertical urban structures can produce strong returns by reflecting signals between buildings and the ground.
• SAR uses side-looking geometry, which can produce distortions such as layover, foreshortening, and shadow.

SAR imagery also contains speckle, a multiplicative rather than additive form of noise. Filtering is commonly applied before further image-processing steps. SAR is useful for observing surface roughness and moisture, and some radar bands can penetrate vegetation canopies more effectively than others.

SAR Bands, Polarization, and Missions

The lecture gives these broad examples of SAR-band applications:

• X band: High-resolution imaging, including urban mapping.
• C band: Land-cover, flood, and agricultural mapping.
• S band: Land-cover and agricultural mapping.
• L band: Forest structure and biomass estimation.
• P band: Greater vegetation penetration and some subsurface studies.

Radar waves can be transmitted and received in different polarizations, producing channels such as HH, HV, and VH.

The lecturer highlights Sentinel-1, a C-band SAR mission operated by the European Space Agency as part of the Copernicus program. Its revisit interval has varied over time, approximately between 6 and 12 days, depending on satellite availability. The lecture also mentions different data products: single-look complex (SLC) data include phase information, while ground-range detected (GRD) data do not. Sentinel-1 acquisition modes include Interferometric Wide (IW) and Stripmap; the lecture notes that resolution can differ across spatial directions.

NISAR, a joint ISRO–NASA mission, is also mentioned; it carries both L-band and S-band SAR systems. The lecturer notes that combining Sentinel-1 SAR with Sentinel-2 multispectral imagery is common in Earth-observation research.

Speakers and Sources Mentioned

• Speaker: An unnamed lecturer/instructor; no additional speakers are heard.
• Organizations and programs referenced: NPTEL, IIT Delhi (channel context); European Space Agency (ESA); Copernicus program; ISRO; NASA.
• Missions and sensors referenced: Sentinel-1, Sentinel-2, TerraSAR-X, COSMO-SkyMed, and NISAR.