nnAEU: a Simple and Efficient CNN-Autoencoder for Hyperspectral Unmixing

Edgard Dabier, Christophe Kervazo, Pietro Gori, Florence TupinarXiv 2026 · arXiv

TL;DR

Researchers have developed nnAEU ("Not New Autoencoder Unmixing"), a lightweight CNN-based model that outperforms state-of-the-art complex architectures like Transformers and Mamba-based models in hyperspectral unmixing. By pruning the "fancy" components from existing top-tier models, the authors created a solution that is up to 20 times smaller and significantly faster while setting a new benchmark for accuracy.

What Problem Does This Paper Address?

Hyperspectral images (HSIs) capture hundreds of narrow spectral bands, allowing us to identify materials from space or drones. However, because of low spatial resolution, a single pixel often contains a "mix" of different materials (e.g., grass, soil, and water). Hyperspectral Unmixing (HSU) is the process of decomposing these mixed pixels into their constituent pure spectra (endmembers) and their relative proportions (abundances).

In recent years, the field has seen an explosion in architectural complexity. Researchers have moved from simple autoencoders to Convolutional Neural Networks (CNNs), then to Transformers, and most recently to State-Space Models like Mamba. Each step claims to advance the state-of-the-art, but this "complexity creep" often comes at the cost of reproducibility and massive parameter counts. The authors of this paper ask a critical question: Is this added complexity actually responsible for the reported gains?

Methods

The authors started with UNMamba, a current leader in the field that uses complex Mamba blocks to capture spatial-spectral dependencies. Through a rigorous ablation (pruning) study, they discovered a surprising result: removing the Mamba blocks entirely didn't just maintain performance—it actually improved it.

The resulting architecture, nnAEU, is a simple CNN-autoencoder:

  1. Encoder (Abundance Module): A series of 2D convolutional layers, group normalization, and SiLU nonlinearities that map the input HSI to abundance maps. It uses a softmax layer to ensure abundances are non-negative and sum to one.
  2. Decoder (Endmember Module): A simplified linear layer where the weights represent the endmembers.
  3. LMM Module: The final reconstruction is a simple linear combination of the estimated endmembers and abundances, following the standard Linear Mixture Model (LMM).
Comparison of the complex UNMamba architecture vs. the streamlined nnAEU.

Datasets

The model was evaluated on four standard, widely-used HSU benchmarks:

  • Samson: 95 × 95 pixels, 156 bands, 3 endmembers.
  • Jasper Ridge: 100 × 100 pixels, 198 bands, 4 endmembers.
  • Apex: 110 × 110 pixels, 285 bands, 4 endmembers.
  • Urban: The largest dataset, 307 × 307 pixels, 162 bands, 6 endmembers.

Key Insights and Ideas

  • The "Complexity Trap": In HSU, models are often trained on a single image. The authors argue that heavy, expressive models (like Transformers) are overkill for such limited data. A simple convolutional backbone is often sufficient to capture all the available signal.
  • Ablation as a Design Tool: By systematically stripping away parts of a SOTA model, the authors found that the "revolutionary" components were actually dead weight.
  • Reproducibility Crisis: The paper highlights a major issue in the field: many published SOTA results cannot be replicated because the code is missing or relies on undocumented "tricks." The authors provide minimal, reproducible code to set a better standard.

Key Results

The quantitative results are striking. As shown in the comparison table below, nnAEU achieves the #1 Rank overall across all datasets.

Performance comparison across four datasets. Note the 'Size' column showing nnAEU's tiny footprint.
  • Efficiency: On the Urban dataset, nnAEU uses only 16k parameters, compared to UNMamba’s 3.8 million. That is a reduction of over 200x.
  • Speed: Training times are significantly lower. For example, on the Urban dataset, nnAEU takes 4 seconds compared to UNMamba's 202 seconds.
  • Accuracy: Despite its simplicity, nnAEU consistently achieves lower Spectral Angular Distance (SAD) and Root Mean Squared Error (RMSE) than its more complex rivals.

Qualitatively, the abundance maps produced by nnAEU (shown in bold in the figure below) are cleaner and more closely match the ground truth (GT) than many complex predecessors.

Fig. 2. Comparison of all method’s estimated bA on the Apex dataset, our proposed method is in bold.

Limitations and Future Work

The authors acknowledge two main limitations:

  1. Standard Benchmarks: Their evaluation is restricted to standard, relatively small benchmark scenes. It is possible that Mamba or Transformer components could prove their worth on much larger, multi-image datasets or "Foundation Models" for remote sensing.
  2. Linear Assumption: The model relies on the Linear Mixture Model (LMM). While widely used, it doesn't account for non-linear mixing effects (like light bouncing between a tree and the ground), which remains a challenge for future simple architectures.

In conclusion, nnAEU serves as a powerful reminder for the AI community: before adding the next "attention" or "state-space" layer, check if a simple convolution can do the job better.

Summary updated

Abstract

Deep learning-based hyperspectral unmixing (HSU) has followed a trajectory of increasing architectural complexity, from simple autoencoders to transformers and, most recently, Mamba-based models, each claiming to advance the state of the art. Contrary to this trend, we propose a lightweight CNN-based autoencoder, coined "Not New Autoencoder Unmixing" (nnAEU). It is derived from the backbone of UNMamba (2025) through a pruning study that removes the Mamba component entirely. nnAEU uses less than 10 times fewer parameters than UNMamba, and far fewer than other transformer- and GAN-based methods, while outperforming them in accuracy. We provide a complete, minimal, reproducible code, available at https://gitlab.telecom-paris.fr/ring/nnAEU.git.
BibTeX
@misc{dabier2026nnaeu,
  author = {Edgard Dabier and Christophe Kervazo and Pietro Gori and Florence Tupin},
  title = {{nnAEU: a Simple and Efficient CNN-Autoencoder for Hyperspectral Unmixing}},
  year = {2026},
  eprint = {2611.17956},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17956},
}
Figure