Dunes Identification Based on Attention Mechanism With Dual-Branch Codec
The research on recognition of dune forms is of great significance for mastering desert landforms and controlling them. We first reviews the status quo of desertification monitoring and control in desert remote sensing image classification and dune form type recognition, and further expounds the adv...
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/10962321/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The research on recognition of dune forms is of great significance for mastering desert landforms and controlling them. We first reviews the status quo of desertification monitoring and control in desert remote sensing image classification and dune form type recognition, and further expounds the advantages and significance of combining remote sensing image and deep learning with dune form type recognition. However, there are still some shortcomings in deep learning-based dune form type recognition, including lack of data set, poor network adaptation, and low segmentation accuracy. Thus, we takes Tengger Desert as the research area and extracts and identifies its dune form types based on deep learning. Specific work contents are as follows: A dual-branch codec dune morphological type segmentation model based on attention mechanism is proposed. In the dual-branch structure, the overcomplete and incomplete networks can take into account both small and large receptive fields, improving the situation of local detail loss caused by the incomplete network in the traditional semantic segmentation structure. The codec hybrid module makes the deep global information interact with the shallow detail information in the dual-branch network to obtain richer feature information. The multiscale mixed attention module is used to extract deep features, and lightweight upsampling operator is used to achieve feature recombination and reduce the number of network parameters. A series of ablation experiments, effectiveness analyses, and comparative studies across different algorithms on two datasets were conducted to evaluate the generalization ability of dual-branch codec network based on attention mechanism across diverse datasets. Using metrics such as Pixel Accuracy, F1-score, and mean intersection over union, its superior recognition performance among various algorithms was validated. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |