A New SCAE-MT Classification Model for Hyperspectral Remote Sensing Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral remote sensing images (HRSI) present challenges in classification due to spectral similarities and manual labeling difficulties, creating "small sample" datasets.
• Deep neural networks require substantial labeled data for effective feature extraction and classification accuracy in HRSI.

Purpose of the Study:

• To propose a novel stacked convolutional autoencoder network model transfer (SCAE-MT) for improved hyperspectral remote sensing image classification.
• To address the "small sample" problem in HRSI classification by leveraging transfer learning with deep feature extraction.

Main Methods:

• Utilized a stacked convolutional autoencoder network to extract deep features from HRSI.
• Applied a transfer learning strategy to develop the SCAE-MT model for scenarios with limited training samples.
• Evaluated the SCAE-MT method on two HRSI datasets, comparing its performance against CAE and SCAE models using varying training set percentages (5%, 10%, 15%).

Main Results:

• The SCAE-MT method demonstrated superior classification performance compared to CAE and SCAE models across all tested training set sizes.
• Achieved an average improvement in overall accuracy (OA) of 2.71%, 3.33%, and 3.07% over CAE and SCAE with 5%, 10%, and 15% training datasets, respectively.
• The proposed method effectively handles the challenge of small sample sizes in HRSI classification.

Conclusions:

• The SCAE-MT method offers a robust solution for hyperspectral remote sensing image classification, particularly in low-data regimes.
• The integration of stacked convolutional autoencoders and transfer learning significantly enhances classification accuracy for HRSI.
• The study validates the effectiveness and superiority of the SCAE-MT approach for HRSI analysis.