[Paper Review] Band Selection from Hyperspectral Images Using Attention-based Convolutional Neural Networks

Pablo Ribalta Lorenzo, Łukasz Tulczyjew+2 more|arXiv (Cornell University)|Oct 24, 2018

Remote-Sensing Image Classification21 references20 citations

TL;DR

This paper proposes an attention-based convolutional neural network (CNN) for hyperspectral image (HSI) band selection that embeds band importance learning within the training process, using attention mechanisms to identify informative bands and anomaly detection to extract a minimal, high-accuracy subset. The method achieves state-of-the-art classification performance using only 9–27% of the original bands, significantly reducing data volume and training time without sacrificing accuracy.

ABSTRACT

This paper introduces new attention-based convolutional neural networks for selecting bands from hyperspectral images. The proposed approach re-uses convolutional activations at different depths, identifying the most informative regions of the spectrum with the help of gating mechanisms. Our attention techniques are modular and easy to implement, and they can be seamlessly trained end-to-end using gradient descent. Our rigorous experiments showed that deep models equipped with the attention mechanism deliver high-quality classification, and repeatedly identify significant bands in the training data, permitting the creation of refined and extremely compact sets that retain the most meaningful features.

Motivation & Objective

• Address the challenge of high-dimensional hyperspectral data in Earth observation, where transfer, storage, and processing costs are prohibitive.
• Overcome limitations of traditional filter and wrapper band selection methods, which are either non-interactive with learning or computationally expensive.
• Develop an embedded band selection approach that learns informative bands during CNN training, avoiding the need for iterative optimization.
• Enable efficient, real-time, and cost-effective HSI data compression suitable for on-board satellite processing.
• Improve interpretability by identifying which spectral bands contribute most to classification performance.

Proposed method

• Integrate a painless attention mechanism into a CNN architecture to learn spatial and spectral attention maps during end-to-end training.
• Use the attention maps to quantify the importance of each spectral band across the input HSI.
• Apply an anomaly detection algorithm (e.g., isolation forest) to the attention map values to identify the most informative bands as outliers.
• Select only a small subset of bands (e.g., 14–19% for Salinas, 9–27% for Pavia) that correspond to high-attention values.
• Train the CNN on the selected bands to evaluate classification performance, ensuring the model remains accurate and efficient.
• Ensure the method is architecture-agnostic and applicable to any HSI dataset and CNN backbone.

Experimental results

Research questions

• RQ1 Can attention mechanisms embedded within a CNN effectively identify the most informative spectral bands in hyperspectral images during training?
• RQ2 Does the proposed method outperform existing filter and wrapper band selection techniques in terms of classification accuracy and computational efficiency?
• RQ3 To what extent can the number of bands be reduced while preserving classification performance?
• RQ4 How robust is the method across different HSI datasets with varying spectral characteristics?
• RQ5 Can the attention maps enhance interpretability by revealing which spectral regions are most relevant for classification?

Key findings

• The attention-based CNN achieves state-of-the-art classification accuracy on both Salinas Valley and Pavia University datasets using only 14–19% of the original bands.
• The method reduces the number of bands to 9–27% of the original in the Pavia University dataset while maintaining high classification performance.
• The inclusion of attention modules does not degrade classification accuracy or increase training time compared to standard CNNs.
• Anomaly detection on attention maps successfully isolates the most informative bands, which are consistently located in specific spectral regions across datasets.
• The approach is significantly faster than traditional wrapper methods, which require iterative training and evaluation of multiple band subsets.
• The method is data-driven and generalizable, applicable to any HSI dataset and CNN architecture without architectural modification.