HyperHazeOff: Hyperspectral Remote Sensing Image Dehazing Benchmark.

Area of Science:

• Remote Sensing
• Image Processing
• Environmental Science

Background:

• Hyperspectral remote sensing images (HSIs) are crucial for environmental and agricultural monitoring.
• Atmospheric haze degrades HSIs, distorting spatial and spectral information and hindering analysis.
• Existing hyperspectral dehazing research lacks paired real-haze benchmarks, limiting fair evaluation and generalization.

Purpose of the Study:

• To introduce HyperHazeOff, the first comprehensive benchmark for hyperspectral dehazing.
• To provide a unified platform for data, tasks, and evaluation protocols in hyperspectral dehazing research.
• To facilitate reproducible research and advance the field of hyperspectral image restoration.

Main Methods:

• The benchmark includes RRealHyperPDID (110 paired real-haze/haze-free HSI scenes) and RSyntHyperPDID (2616 synthetic paired samples).
• Physically grounded haze formation models were used for synthetic data generation.
• Annotations for agricultural field delineation and land classification are provided for downstream task assessment.

Main Results:

• State-of-the-art hyperspectral models trained on existing datasets fail to generalize to real-world haze.
• Training on the synthetic dataset (RSyntHyperPDID) within HyperHazeOff significantly improves real-haze restoration by AACNet.
• HyperHazeOff establishes reproducible baselines for hyperspectral dehazing.

Conclusions:

• HyperHazeOff is essential for advancing hyperspectral dehazing research due to its comprehensive nature and real-world data.
• The benchmark demonstrates the limitations of current models and highlights the importance of realistic training data.
• The open availability of HyperHazeOff promotes reproducible research and facilitates the development of more robust hyperspectral dehazing techniques.