Scholarlink

<p>This study evaluated the impact of optimized channel selection and density-based noise reduction on the precision of land classification in SAR multispectral image fusion. Remote sensing technologies play a crucial role in land cover mapping and environmental monitoring studies. However, the effectiveness of classification processes depends heavily on the appropriate fusion of SAR and multispectral images and the selection of the most suitable spectral channels. Using all multispectral channels directly during fusion leads to data redundancy and higher computational costs. This study evaluates the impact of optimized channel selection and noise reduction based on density on the accuracy of land classification in SAR-multispectral image fusion. The analyzes were performed using a dataset consisting of Sentinel-1 SAR and Sentinel-2 multispectral images. Rather than using all multispectral bands directly, the Particle Swarm Optimization (PSO) algorithm was used to determine the most suitable channels. Five metrics (ERGAS, entropy, quality index (Q), RASE, and RMSE) were tested to evaluate fusion quality, with ERGAS being identified as the most effective channel selection criterion. The selected channels were then merged with SAR data using the NSST-MSMG-PCNN and MWGF merging algorithms. Classification was performed using CNN-based deep learning models, including ResNet100, YOLOv8, and DenseNet121. The experimental results demonstrate that channel selection significantly improves classification accuracy. The precision increased from 68.2% to 73.3% for the NSST-MSMG-PCNN method and from 63.4% to 67.5% for the MWGF method. The DenseNet121 model achieved the highest accuracy rate. In conclusion, combining optimized channel selection with density-based denoising methods significantly improves fusion quality and classification performance. The proposed approach offers a computationally efficient and scalable method to improve the precision of land classification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating the Effect of Channel Selection on Classification Success in SAR-Multispectral Fusion

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Semih Gencay,
  • Caner Ozcan

摘要

This study evaluated the impact of optimized channel selection and density-based noise reduction on the precision of land classification in SAR multispectral image fusion. Remote sensing technologies play a crucial role in land cover mapping and environmental monitoring studies. However, the effectiveness of classification processes depends heavily on the appropriate fusion of SAR and multispectral images and the selection of the most suitable spectral channels. Using all multispectral channels directly during fusion leads to data redundancy and higher computational costs. This study evaluates the impact of optimized channel selection and noise reduction based on density on the accuracy of land classification in SAR-multispectral image fusion. The analyzes were performed using a dataset consisting of Sentinel-1 SAR and Sentinel-2 multispectral images. Rather than using all multispectral bands directly, the Particle Swarm Optimization (PSO) algorithm was used to determine the most suitable channels. Five metrics (ERGAS, entropy, quality index (Q), RASE, and RMSE) were tested to evaluate fusion quality, with ERGAS being identified as the most effective channel selection criterion. The selected channels were then merged with SAR data using the NSST-MSMG-PCNN and MWGF merging algorithms. Classification was performed using CNN-based deep learning models, including ResNet100, YOLOv8, and DenseNet121. The experimental results demonstrate that channel selection significantly improves classification accuracy. The precision increased from 68.2% to 73.3% for the NSST-MSMG-PCNN method and from 63.4% to 67.5% for the MWGF method. The DenseNet121 model achieved the highest accuracy rate. In conclusion, combining optimized channel selection with density-based denoising methods significantly improves fusion quality and classification performance. The proposed approach offers a computationally efficient and scalable method to improve the precision of land classification.