a novel region-specific deep image fusion framework via pso clustering and lsqr optimization

This research proposes a novel hybrid framework for satellite image fusion, combining unsupervised clustering, deep convolutional learning, and sparse numerical optimization. classical pansharpening methods such as ihs, pca, and dwt, although computationally lightweight, exhibit poor adaptability to scene heterogeneity and often introduce spectral distortion. deep learning models like cnns and resnets offer powerful spatial-spectral representation capabilities but struggle with overgeneralization when applied globally. our model addresses these limitations by first segmenting the scene using particle swarm optimization (pso), then applying region-specific cnn or resnet fusion, and finally harmonizing global consistency using lsqr optimization. experiments on landsat 8 and spot-6 datasets confirm superior performance across rmse, uiqi, cc, and ergas metrics, validating the model’s robustness and practical potential.