A Novel Multiscale Feature Fusion With Adaptive Scale-Space Pyramid Network for Aerial Scene Recognition Using Remote Sensing Images

Remote Sensing is an area anthropogenic study undertaken worldwide. It has succeeded significantly in important applications such as climate monitoring, disaster prediction and land use planning. However, due to the diversity of scales, intra-class similarities, and complex scenes, the accurate recognition process remains challenging. Transformers' global attention mechanism helps them to overcome the limitations of CNNs' local receptive fields; however, they have drawback of increased computing complexity. To overcome such challenges, this work proposes an Adaptive Scale-Space Pyramid Network (ASSPN) for improved remote sensing image classification. The ASSPN architecture contains a learnable Gaussian pyramid module for multi-scale feature representation, a scale selection attention mechanism for dynamically weighing feature relevance, a cross-feature propagation module for fusion guided by uncertainty, and a complexity-aware adaptive pooling module for preserving semantic discriminative features. Experiments are performed three benchmark datasets such as EuroSAT, NWPU-RESISC-45, and MLRSNet. On these datasets, the ASSPN achieves state-of-the-art results with accuracies of 96.14%, 94.73%, and 95.42%, respectively. The obtained accuracy is outperforming previous CNN and transformer-based systems with significant margins. Furthermore, ASSPN is noise perturbation-resistant and shows generalization capability across a wide range of land-cover categories. Ablation studies established the complementary benefits of the core modules, while LIME-based explainability analysis confirmed the predicative trustworthiness of the model.