QUH-Qingyun HSI Benchmark

• QUH-Qingyun is a hyperspectral imaging dataset benchmark designed for spectral–spatial classification using PCA-reduced 15 band patches.
• It employs 9×9 patch extraction with a 1%/1%/98% train/validation/test split to simulate extreme low-supervision conditions.
• Performance metrics, including 93.86% OA, 86.83% AA, and a 0.9186 Kappa, validate its effectiveness for evaluating deep learning models.

The QUH-Qingyun dataset is a remote sensing hyperspectral image (HSI) benchmark introduced and utilized in the context of spectral–spatial classification research. It serves as one of two evaluation platforms—the other being QUH-Tangdaowan—for the assessment of advanced deep learning models such as SS-MixNet in highly label-sparse regimes, specifically 1% supervised training settings. The dataset constitutes a critical reference point for constructing, benchmarking, and comparing lightweight networks aimed at robust HSI scene interpretation in constrained annotation scenarios (Alkhatib, 19 Nov 2025).

1. Dataset Characteristics and Structure

The QUH-Qingyun dataset is explicitly described as an HSI classification dataset with the following properties:

• Number of Classes: 6
• Spectral Bands: After principal component analysis (PCA), each patch contains $P = 15$ bands. This is a reduced dimensionality from the raw channel count.
• Spatial Dimensions: Patches are extracted as $9 \times 9$ pixel windows across the image.
• Patch Format: Each input patch after preprocessing has shape $X \in \mathbb{R}^{9 \times 9 \times 15}$, suitable for direct input to spectral–spatial deep learning models.
• Split: The split protocol utilizes 1% of pixels for training, 1% for validation, and 98% for testing, with random sampling performed within each semantic class to support generalized evaluation with minimal supervision.

This patch-based formulation enables systematic study of spectral–spatial relationships and efficient training of deep neural networks under stringent annotation budgets.

2. Preprocessing and Dimensionality Reduction

HSI patches undergo an initial dimensionality reduction step via PCA:

• Principal Component Analysis: The raw spectral dimension ($C$ bands) is projected onto the top $P=15$ principal components, matching the dimensionality used for QUH-Tangdaowan and ensuring computational tractability with strong information preservation.
• Patch Extraction: A sliding window of $9 \times 9$ is used to partition the scene into overlapping or non-overlapping samples for model input.
• These preprocessing steps are implemented in the data_loader.py component of the official code release for reproducibility (Alkhatib, 19 Nov 2025).

3. Benchmark Usage and Protocols

The QUH-Qingyun dataset is employed as a standard testbed for spectral–spatial HSI methods, with the protocol aiming to mirror realistic label scarcity:

• Training Regime: Only 1% of annotated pixels per class are used for both training and validation phases, simulating extreme low-supervision scenarios.
• Model Input: Each sample fed to the neural model preserves both the low-dimensional spectral ($P$ bands) and local spatial ($9 \times 9$ region) context.
• Comparison: Performance on QUH-Qingyun is routinely benchmarked alongside QUH-Tangdaowan to validate generalization and robustness across scenes of differing class granularity.

4. Performance Metrics and Results

The following metrics are applied to measure discriminative capacity and generalization:

Metric	Definition
OA (Overall Accuracy)	Pixel-wise proportion of correct predictions over the entire test set
AA (Average Accuracy)	Mean per-class accuracy
Kappa	Cohen’s Kappa coefficient, measuring class-wise agreement beyond random chance

In the pivotal study introducing SS-MixNet, the following results were obtained on QUH-Qingyun (Table 2, (Alkhatib, 19 Nov 2025)):

• OA: 93.86%
• AA: 86.83%
• Kappa: 0.9186

Per-class accuracy was highest among compared models for "Trees" (95.60%), "Concrete building" (94.63%), and "Asphalt road" (92.55%), demonstrating the dataset's suitability for evaluating nuanced intra-scene distinctions.

5. Comparative Methods and Ablation Studies

QUH-Qingyun serves as an evaluation domain for a suite of spectral–spatial classifiers, including:

• 2D-CNN
• 3D-CNN
• IP-SWIN
• SimPoolFormer
• HybridKAN

SS-MixNet outperformed these baselines in overall and per-class accuracy under identical 1% label regimes. Ablation on related datasets shows incremental gains from spectral and spatial mixers as well as from depthwise attention, underscoring the importance of retaining joint local and contextual spectral–spatial cues.

6. Implementation and Reproducibility

All code to preprocess, train, and evaluate models on QUH-Qingyun is publicly available at [https://github.com/mqalkhatib/SS-MixNet]:

• Components:
  • data_loader.py (PCA, patch extraction, splits)
  • model.py (network definitions)
  • train.py (optimization and early stopping)
  • evaluate.py (metric calculation and map generation)
• Reproducibility: By following the documented pipeline and split protocol, researchers can reproduce all benchmark tables, figures, and ablation studies as reported.

7. Significance and Research Impact

The QUH-Qingyun dataset establishes a rigorous benchmark for label-efficient HSI scene classification. Its integration into the SS-MixNet evaluation framework, alongside multiple established baselines, facilitates objective, reproducible assessment of emerging spectral–spatial architectures under challenging data regimes. The dataset's PCA-compressed, patch-based format supports both scalable experimentation and exploration of local/global spectral–spatial feature learning paradigms (Alkhatib, 19 Nov 2025).