DnRCNN: Deep Recurrent Convolutional Neural Network for HSI Destriping.

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Hyperspectral image (HSI) restoration methods often suffer from spectral and spatial information loss.
• Existing deep learning approaches neglect the inherent correlations within HSI data, limiting restoration quality.

Purpose of the Study:

• To propose an innovative deep recurrent convolutional neural network (DnRCNN) for hyperspectral image (HSI) destriping.
• To explore HSI destriping by leveraging inner band and interband correlations using recurrent convolutional neural networks.

Main Methods:

• Developed a novel DnRCNN model incorporating a selective recurrent memory unit (SRMU).
• SRMU is designed to extract correlative features from spectral and spatial domains independently.
• Implemented a recurrent fusion (RF) strategy with group concatenation to integrate complementary features for noise removal and detail preservation.

Main Results:

• The proposed DnRCNN model effectively removes strip noise from HSI data.
• The method successfully preserves crucial scene details during the destriping process.
• Extensive experiments demonstrate that the DnRCNN achieves new state-of-the-art (SOTA) performance in HSI destriping.

Conclusions:

• The DnRCNN model offers a significant advancement in hyperspectral image destriping.
• Exploiting spectral and spatial correlations via recurrent networks is crucial for effective HSI restoration.
• The proposed SRMU and RF strategies contribute to superior noise removal and detail preservation.