LECTURE 21

Summary

The lecture introduces synthetic aperture radar (SAR) and explains how AI and deep learning can be used to analyze SAR imagery. It sets up the week’s focus on SAR applications, including object detection, despeckling, change detection, and architectures designed to use SAR-specific information.

What Makes SAR Different

• SAR is an active remote-sensing system: It transmits microwave signals and records the energy scattered back from Earth. A SAR pixel represents radar backscatter, not reflected sunlight.
• Unlike optical sensors, SAR can acquire images at night and through cloud cover, making it valuable during cloudy seasons and emergencies when usable optical imagery is unavailable.
• Depending on the radar band, SAR signals can penetrate vegetation to varying degrees, supporting applications such as forest and biomass analysis.
• SAR imagery can be difficult to interpret visually. Object appearance depends strongly on imaging geometry and physical properties such as surface roughness, moisture, dielectric properties, wavelength, and polarization.
• SAR’s side-looking geometry creates distinctive effects and distortions, including layover, foreshortening, and shadow. The lecture also treats range and azimuth resolution separately.
• Backscatter patterns provide useful clues:
  • Smooth surfaces, such as calm water, tend to reflect energy away from the sensor and appear dark.
  • Rough surfaces often scatter energy in multiple directions and appear brighter.
  • Urban structures can produce complex, bright returns.
• Speckle gives SAR images a grainy appearance. It is multiplicative rather than additive noise, and despeckling is often performed as preprocessing.

Bands, Polarization, and Missions

Radar bands have different characteristics and suit different applications:

• X-band: High-resolution imaging, including urban mapping. Examples include TerraSAR-X and COSMO-SkyMed.
• C-band: Land-cover, flood, and agricultural mapping. Sentinel-1 is a widely used, freely available C-band mission.
• S-band: Land-cover and agricultural applications.
• L-band: Forest structure and biomass estimation.
• P-band: Greater vegetation penetration and potential for subsurface studies.

Polarization also matters. Different land-cover types interact differently with radar waves transmitted or received in different polarizations, providing additional information for analysis.

The lecture highlights Sentinel-1’s availability and several acquisition modes. It distinguishes Single Look Complex (SLC) data, which includes phase information, from Ground Range Detected (GRD) data, which does not. It also introduces NISAR, a joint NASA–ISRO mission carrying both L- and S-band SAR systems.

Deep-Learning Opportunities and Challenges

Deep learning can learn features directly from SAR data, reducing reliance on handcrafted texture descriptors and enabling end-to-end models. This is especially helpful because SAR images are often difficult to interpret visually.

Important challenges include:

• Limited labeled data, particularly for dense tasks such as semantic segmentation and object detection.
• Differences between SAR datasets, including variations in band and polarization, which make domain adaptation and generalization important.
• The difficulty of transferring models trained on optical imagery. Optical pre-trained models generally cannot simply be applied to SAR, although fine-tuning may work in some settings; training from scratch is another option.
• Deciding whether to use standard architectures or adapt them to SAR-specific information, such as phase or viewing angle.

Common SAR Deep-Learning Tasks

The lecture surveys several task types:

• Despeckling: Often used as preprocessing and performed with traditional filters or deep-learning methods.
• Image classification
• Semantic segmentation and other dense prediction
• Regression: For example, biomass estimation. The input data, output layer, and loss function must suit the regression task.
• Object detection: A major SAR research area, particularly for targets such as ships; building detection is another example.
• Change detection: Includes disaster monitoring and combinations of SAR with optical imagery.

For supervised SAR change detection, the lecture outlines a general pipeline:

1. Extract features from the pre-change and post-change images.
2. Compare the extracted features.
3. Fuse the features.
4. Produce pixel-wise change predictions.

SAR can be especially useful when optical imagery is unavailable after an event. Pre-event and post-event observations may also come from different sensors.

Research Directions

The lecture identifies several recurring themes in SAR and deep-learning research:

• Using labeled optical data to help train SAR models, for example through feature knowledge distillation or domain adaptation.
• Transferring knowledge between SAR datasets, such as using labeled data from one band to support a model for another band.
• Combining SAR and optical imagery so that each sensor can complement the other.
• Designing SAR-specific architectures that exploit phase, viewing angle, or other radar properties.
• Generative approaches, including generating SAR imagery or translating between SAR and optical imagery.
• Explainability, uncertainty, and physics-guided models, which can help interpret predictions from imagery that is not visually intuitive.

Feature Distillation for SAR Segmentation

One possible approach to addressing the scarcity of SAR segmentation labels is feature distillation:

1. Train a model on optical imagery.
2. Freeze the optical model.
3. Train a SAR model using available SAR segmentation labels.
4. Add a distribution-based loss that encourages the SAR model to produce features similar to those of the optical model.

The aim is to transfer useful feature knowledge while training the SAR model for its target task.

Speakers and Sources

• Speaker: One lecturer, not named in the subtitles, presenting the NPTEL course AI for Earth Observation.
• Sources or missions mentioned: Sentinel-1 (European Space Agency), TerraSAR-X (DLR), COSMO-SkyMed (Italy), and NISAR (NASA and ISRO).