Mapping Peatlands Combing Deep Learning With Sparse Spectral Unmixing Based on Zhuhai-1 Hyperspectral Images
The mixed pixel problem, arising from the complex vegetation types of peatlands, poses a significant challenge for remote sensing-based peatland mapping. A convolution and transformer-based reconstruction and sparse unmixing algorithm that integrates deep learning and sparse spectral unmixing is pro...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11072024/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The mixed pixel problem, arising from the complex vegetation types of peatlands, poses a significant challenge for remote sensing-based peatland mapping. A convolution and transformer-based reconstruction and sparse unmixing algorithm that integrates deep learning and sparse spectral unmixing is proposed to address the spectral variability and spatial heterogeneity of the endmembers in hyperspectral datasets. Dimensionality reduction was performed on the hyperspectral data using convolution and pooling operations in a convolutional neural network. A multihead attention mechanism based on the transformer encoder captures global contextual information by correlating patches in the image, facilitating precise hyperspectral data reconstruction. Sparse spectral unmixing was applied to the reconstructed global features to obtain classification maps. Experiments were conducted on three representative peatlands in China: the Honghe, Zhalong, and Nanwenghe Nature Reserves. The proposed method outperformed the fully constrained least squares, sparsely constrained least squares spectral unmixing algorithms and deep transformer network hyperspectral unmixing, achieving a mean root-mean-square error (RMSE) of 0.1548 to 0.1781 and a mean spectral angle distance (SAD) of 0.1101 to 0.2211. The greatest improvements were 28.2% in the RMSE and 25.4% in the SAD, demonstrating the robustness and generalization capability of the algorithm. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |