Decoupled Complementary Spectral–Spatial Learning for Background Representation Enhancement in Hyperspectral Anomaly Detection

Authors: Wenping Jin, Yuyang Tang, Li Zhu, Fei Guo

Overview

This repository provides the official implementation for the training pipeline and the final fusion-based anomaly detection framework introduced in our paper pdf :

"Decoupled Complementary Spectral–Spatial Learning for Background Representation Enhancement in Hyperspectral Anomaly Detection"

Our method introduces a complementary learning paradigm that enhances background spectral and spatial representations through a novel rebellious student mechanism. The framework is composed of two sequential training stages, as illustrated below.

Stage 1 — Spectral Enhancement via Reverse Distillation

We employ a reverse-distillation-based pruning strategy to obtain a compact network that focuses on enhancing background spectral features.

Stage 2 — Training the Spatial–Spectral Rebellious Student

Using the spectral enhancement network (FEN) as a teacher, we train a spatial–spectral network that intentionally learns features complementary to the teacher, rather than imitating it.

HAD100 Dataset

HAD100: https://zhaoxuli123.github.io/HAD100/

Obtaining Detection Results on HAD100

To obtain anomaly detection results on the HAD100 dataset, execute the following commands. You can specify the number of spectral bands to use with the --input_channel argument:

• For first 50 bands test:

python test.py --input_channel 50

• For first 100 bands test:

python test.py --input_channel 100

• For first 200 bands test:

python test.py --input_channel 200

Training on HAD100

To train your own model on the HAD100 dataset, run the following command:

step 1:

cd train_FERD
python train.py

step 2:

python train.py

Anomaly Detection Results on HAD100

The following table presents the mean AUC (mAUC) results achieved using different numbers of spectral bands:

Bands	mAUC
The First 50 Bands	0.9953
The First 100 Bands	0.9910
The First 200 Bands	0.9902

Obtaining Detection Results on ABU

The corresponding code is located in the FERS_ABU directory.

Using the RX detector:

python test.py

Using the AE detector:

python test.py --detect AE

Citation

If this repo works positively for your research, please consider citing our paper. Thanks all!

@misc{jin2025_FERS,
      title={Rebellious Student: A Complementary Learning Framework for Background Feature Enhancement in Hyperspectral Anomaly Detection}, 
      author={Wenping Jin and Yuyang Tang and Li Zhu and Fei Guo},
      year={2025},
      eprint={2510.18781},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.18781}, 
}

@article{jin2024_FERD,
  author={Jin, Wenping and Dang, Feng and Zhu, Li},
  journal={IEEE Geoscience and Remote Sensing Letters}, 
  title={Feature Enhancement with Reverse Distillation for Hyperspectral Anomaly Detection}, 
  year={2024},
  doi={10.1109/LGRS.2024.3456178}
}