Citation

Kibria, Nafisa Binte Ghulam and Kibria, Hafsa Binte and Morol, Md Kishor and Liew, Tze Hui and Nandi, Dip (2026) Dual-stream deep learning with physics-informed attention for multispectral satellite image classification. Remote Sensing Applications: Society and Environment, 43. p. 102082. ISSN 23529385

Text 8.pdf - Published Version Restricted to Repository staff only Download (2MB)

Abstract

Multispectral satellite imagery provides rich spectral information that is critical for accurate land use and land cover (LULC) classification. However, most deep learning approaches either rely solely on RGB bands or process all spectral bands uniformly, ignoring their distinct physical properties and increasing computational cost. In this work, a lightweight dual-stream deep learning framework has been proposed that explicitly exploits complementary RGB and multispectral information using physics-informed attention mechanisms. This architecture processes RGB bands using a pretrained MobileNetV2 backbone, while selected spectral bands (Red Edge, Near-Infrared, and Short-Wave Infrared) are processed through a custom lightweight convolutional network. To enhance spectral feature learning, a physics-informed band attention module is introduced, incorporating wavelength-aware weighting and inter-band correlation modeling. Multiple fusion strategies, including early, mid-level, and late fusion, are systematically evaluated. Experiments on the EuroSAT dataset demonstrate that the proposed model achieves a classification accuracy of 98.82% ± 0.08% with only 3.86M parameters, while achieving competitive accuracy with superior model efficiency compared to existing state-of-the-art approaches. This improvement is achieved by explicitly exploiting complementary RGB and nonvisible spectral information through a physics-informed attention mechanism, enabling more discriminative feature learning. In addition, the proposed model provides a favorable accuracy– efficiency trade-off, achieving faster inference (64.97 ± 0.23 ms) compared to conventional attention mechanisms such as CBAM while maintaining competitive or superior performance. Extensive ablation studies further confirm the effectiveness of dual-stream learning, informed band selection, and the proposed attention design.

Item Type:	Article
Uncontrolled Keywords:	Multispectral satellite images, convolutional neural networks
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5101-6720 Telecommunication. Including telegraphy, telephone, radio, radar, television
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	30 Jun 2026 04:57
Last Modified:	30 Jun 2026 04:57
URII:	http://shdl.mmu.edu.my/id/eprint/16138

Downloads

Edit (login required)