ODMAF: Optimization-Driven Multi-Scale Attention Fusion for Multispectral Image Feature Extraction and Recognition

Multispectral image analysis plays a pivotal role in remote sensing, precision agriculture, and environmental monitoring. However, the diverse spatial and spectral characteristics of multispectral data pose significant challenges for robust feature extraction and accurate recognition. In this work, we propose Optimization-Driven Multi-Scale Attention Fusion (ODMAF), a novel framework that integrates multi-scale convolutional feature extraction with an optimization-driven attention fusion mechanism. ODMAF first generates a hierarchy of spatial–spectral features via parallel multi-scale convolutional branches and spectral attention modules. These features are then adaptively fused through a mathematically principled attention mechanism, formulated as a constrained optimization problem with entropy regularization to encourage diversity and interpretability. Unlike prior attention-based methods, ODMAF frames attention weight learning as a constrained, entropy-regularized optimization problem with theoretical guarantees on optimality and generalization, and explicitly unifies spatial and spectral feature hierarchies for interpretable fusion. Comprehensive experiments on four benchmark hyperspectral datasets demonstrate that ODMAF achieves state-of-the-art performance, improving overall accuracy by 1.5–2.3% over the best-performing baselines (e.g. achieving 97.7% OA on Indian Pines and 99.8% on Pavia University) and surpassing several competitive deep learning baselines in terms of overall accuracy, convergence stability, and model interpretability. Visualization and theoretical analysis further verify that ODMAF effectively captures class-specific, multi-scale feature contributions, providing a robust and transparent solution for multispectral image understanding. Beyond benchmark accuracy, ODMAF’s optimization-driven fusion is efficient and scalable, enabling practical deployment in operational remote sensing, agriculture monitoring, and large-scale environmental analytics.