LECTURE 29

Summary

The lecture introduces dimensionality reduction for hyperspectral images and begins a discussion of hyperspectral unmixing.

Why reduce dimensionality?

Hyperspectral images contain hundreds of spectral bands. This provides rich information but also creates several challenges:

• Neighboring bands are often highly correlated, resulting in redundant information.
• High dimensionality increases computational and storage requirements.
• More dimensions generally require more training data. Because annotated hyperspectral data can be scarce, classifiers may overfit when data are limited.
• High-dimensional data can be difficult to interpret and visualize.

Reducing dimensionality can lessen redundancy and improve computational efficiency. It may also benefit classification, interpretation, and visualization. The appropriate degree of reduction depends on the task.

Two approaches to dimensionality reduction

• Feature or band selection: Choose a subset of the original bands—for example, retaining 100 of 230 bands. Because the selected bands retain their original identities and wavelengths, this approach generally preserves more physical interpretability.
• Feature extraction: Transform the original data into a smaller set of new features by combining information across bands. This can produce a more compact representation, but the resulting features are not original spectral bands.

The distinction is that selection discards some bands, while extraction creates transformed features. Both approaches reflect the idea that a dataset may contain fewer underlying dimensions of information than its number of measured bands suggests.

Band-selection methods

The lecture outlines several ways to select informative bands:

• Statistical band ranking: Rank bands using criteria such as entropy. This is generally computationally efficient and does not necessarily depend on a particular classifier.
• Mutual-information-based selection: Measure how much information a band provides about the labels. Bands with greater mutual information may be more informative, after which redundant bands can be removed.
• Classifier-guided selection: Evaluate candidate subsets by training a classifier and comparing its accuracy. This can produce strong task-specific performance, but is computationally expensive because many subsets may need to be tested and the classifier retrained for each one.

Density-based band selection

The lecture describes a density-based approach at a high level:

1. Start with a band-similarity matrix representing similarities between pairs of original bands.
2. Calculate a local density for each band from its similarities to other bands. High density indicates that many similar bands surround it; low density indicates that it is relatively isolated.
3. Sort bands by density. For each band other than the densest one, calculate its distance from bands with higher density. A large distance suggests that the band may be a distinct cluster center.
4. Use density and distance to calculate band-importance scores.
5. Rank the scores and select the desired band indices.
6. Assign every original band, including those not selected, to a selected representative band to produce cluster assignments.

The lecturer does not provide the precise importance-score formula, referring interested students to the original paper.

Deep-learning approaches

Deep networks can learn compact representations and model nonlinear structure, potentially improving downstream tasks.

• Autoencoders: An encoder compresses spectral information into a bottleneck representation, and a decoder reconstructs the original spectra. The bottleneck is a reduced feature representation, so this is feature extraction rather than band selection.
• 1D CNNs: These can learn reduced features together with a downstream objective, such as classification, allowing the reduction to be optimized for that task.
• Channel attention: A channel-weighting mechanism can be applied to hyperspectral data. When used immediately after the input, its learned weights can be interpreted as band reweighting. The weights may also inform band selection: rank them and retain strongly weighted bands. This assumes that the data used later come from a distribution similar to the data used to learn the weights.

Evaluating dimensionality reduction

The lecture names several evaluation criteria:

• Downstream classification accuracy: Check whether performance remains acceptable after reduction. A substantial drop may indicate that important information was lost.
• Reconstruction error: Use this when the method aims to reconstruct the original spectra or data.
• Visualization and interpretability: Consider whether the reduced representation helps with understanding or displaying the data.
• Computational efficiency: Assess whether the reduction meaningfully lowers processing costs.

Introduction to hyperspectral unmixing

The lecture then introduces unmixing: decomposing a pixel’s measured spectrum into contributions from its constituent materials. Mixed pixels arise when the spatial resolution is too coarse for each pixel to contain just one material.

• End members are spectra representing relatively pure materials.
• Abundance fractions describe the proportion of each material in a pixel. They are estimated per pixel and are commonly constrained to lie between zero and one and to sum to one, reflecting a physical mixture.
• Accurate end members are important for reliable unmixing.
• Applications include mineral exploration, precision agriculture, and urban-material identification.

The lecture ends with this introductory overview; the discussion of unmixing is to continue in the next lecture.

Speakers and sources

• Speaker: One lecturer, whose name is not stated in the subtitles.
• Source/channel context: NPTEL IIT Delhi.