LECTURE 28

Hyperspectral Anomaly Detection

The lecture introduces hyperspectral anomaly detection, explaining how it differs from target detection, why it is challenging, and how statistical and deep-learning methods can address it.

1. Anomaly Detection Versus Target Detection

• Hyperspectral images contain hundreds of spectral bands, which can reveal distinctive material signatures and subtle deviations.
• Target detection assumes that a known target signature is available.
• Anomaly detection assumes that the objects of interest are unknown. It looks for spectral rarity—pixels that differ from the background—rather than matching a specified target.
• Spectral information alone may not be enough; the spatial context around a pixel can help determine whether it is genuinely unusual.

2. Main Challenges

• No labeled anomaly examples: If labeled examples of a particular object are available, the task becomes closer to target detection.
• Complex backgrounds: Background statistics can vary substantially, especially in heterogeneous urban scenes, making them difficult to model.
• Small anomalies: Anomalous objects may occupy very few pixels and be difficult to distinguish.
• Atmospheric effects, clouds, and noise: These can distort spectral measurements or produce apparent anomalies.
• High dimensionality: Hundreds of bands increase computational cost. Dimensionality reduction can help, but excessive reduction may discard useful information.
• Uneven band quality: Some bands, including water-absorption bands, may be less reliable and are often excluded.
• Natural spectral variation: Materials can vary with seasons and other conditions, creating false alarms.
• False alarms and missed detections: Their relative importance depends on the application. For example, surveillance may prioritize avoiding missed anomalies, even if that means accepting more false alarms.

3. Statistical Anomaly Detection

A basic statistical approach models the background distribution and flags pixels that deviate significantly from it.

• Gaussian background modeling treats the background as a statistical distribution and identifies pixels that are unusually distant from it.
• Reed–Xiaoli (RX) anomaly detector measures a pixel’s Mahalanobis distance from the background statistics.
• PCA-based detection uses principal component analysis to represent dominant background structure; large reconstruction errors can indicate anomalies.

These methods can use either global or local background statistics.

4. Global, Local, and Dual-Window Context

• Global methods estimate statistics from the whole image or a large area.
  • They can work well in relatively homogeneous scenes.
  • They may perform poorly when the scene contains substantial background variation.
• Local methods estimate background statistics within a confined neighborhood, often using a sliding-window approach.
  • They better accommodate local spatial variability and can be more useful in complex scenes.
• Dual-window methods combine both scales:
  • An inner window captures local variation.
  • An outer window captures broader background information.
  • Comparing the two can reveal whether the local region differs from its surrounding context.

5. Deep-Learning Approaches

Deep networks can learn nonlinear spectral and spatial patterns. The lecture emphasizes that many deep-learning approaches adapt ideas already used in statistical methods.

• Autoencoders
  • Train an encoder–decoder to reconstruct normal background spectra or regions.
  • Background pixels are expected to have relatively small reconstruction errors, while anomalies are expected to have larger errors.
  • A dual-window version can use separate autoencoders for inner and outer windows, then compare their reconstruction errors.
  • Small errors in both windows suggest no anomaly; large errors in both may indicate an anomaly. Other combinations of inner and outer errors can provide additional contextual clues.
  • Convolutional autoencoders can capture spatial context more effectively than a basic autoencoder.
• Generative adversarial networks (GANs) can also be adapted to anomaly detection using ideas similar to reconstruction-based autoencoders.
• 3D CNNs jointly process spatial and spectral dimensions to learn spectral–spatial anomaly patterns.

6. A Possible Weakly Supervised Workflow

The lecture outlines a way to combine anomaly detection with human review and supervised learning:

1. Optionally apply dimensionality reduction.
2. Run an existing fast anomaly-detection method over the area of interest.
3. Identify candidate anomalous pixels or objects, initially allowing a relatively high number of false alarms if avoiding missed anomalies is the priority.
4. Apply post-processing, potentially including human review, to refine the candidate labels.
5. Train a supervised classifier, such as a 3D CNN, using the refined examples.
6. Use the trained model to detect similar anomalies in the relevant geography.

This approach can be geography-dependent: a model refined for one environment, such as a desert, may not transfer well to a very different environment, such as an urban area.

7. Managing False Alarms

• False alarms reduce the reliability of an anomaly detector, but missing anomalies may be more costly in some applications.
• A practical strategy is to favor detecting potential anomalies first, then reduce false alarms through post-processing.
• Post-processing may use contextual knowledge or other detection systems—for example, cloud detection or water and land-cover information—to remove detections associated with known non-target causes.

8. Datasets Mentioned

• Airport–Beach–Urban (ABU) dataset: A collection of scenes mainly acquired using the AVIRIS sensor.
• HYDICE dataset: An urban-area image from the United States, described as 80 × 100 pixels. The number of bands depends on preprocessing; many studies use 175 bands. Its ground truth includes 21 anomalous pixels, mainly cars.

The limited size of available datasets reinforces the importance of unsupervised methods and makes fully supervised anomaly detection difficult.

Speakers and Sources

• Speaker: An unnamed lecturer presenting an NPTEL IIT Delhi course lecture.
• Other speakers or sources: None featured in the subtitles.