Adaptive Deep Image Fusion through PSO Segmentation and Sparse LSQR Optimization

This study presents a region-adaptive hybrid fusion framework that integrates unsupervised clustering, deep convolutional modeling, and sparse numerical optimization to enhance multispectral satellite image fusion. Unlike classical pansharpening approaches, which are computationally efficient but inherently global and unable to adequately handle spatial–spectral variability, and purely deep-learning-based models that often overgeneralize across heterogeneous scenes, the proposed method introduces a region-aware strategy to preserve fine structures and spectrally weak areas. By employing Particle Swarm Optimization–based unsupervised segmentation, the framework dynamically partitions the image into spectrally homogeneous regions, enabling localized CNN/ResNet-based fusion that prevents spectral dilution near region boundaries. The independently fused regions are subsequently unified through LSQR-based global optimization, which enforces structural coherence and minimizes reconstruction error without reliance on large labeled datasets. Experimental evaluations on IKONOS datasets demonstrate that the proposed framework consistently outperforms classical and state-of-the-art fusion methods in terms of UIQI, RMSE, CC, and ERGAS metrics, confirming its robustness, adaptability, and effectiveness for high-fidelity multispectral image enhancement in complex remote-sensing environments.

کلیدواژه ها:

Pansharpening ، Deep Neural Networks ، PSO-Based Segmentation ، LSQR Optimization ، Multispectral–Panchromatic Fusion