Multiscale Scene Localization Based on Composite Network for Remote Sensing Imageries: A Case Study on Critical Infrastructure

This paper proposes an innovative framework to enhance remote sensing image reconstruction for damage detection of critical infrastructure. The approach integrates vision transformer architectures, specifically the Shifted Window (SWIN) transformer, with Global Convolutional Neural Network (GCN) and onboarded Spatial Cross-scale Encoder (SCSC) initially for feature semantic segmentation and multiscale scene localization. The framework scales from global scene analysis to specific damage localization, utilizing a chain of GCN for global feature extraction and SCSC with SWIN transformer for fine-grained damage localization at components and assets scales. A case study on critical infrastructure demonstrates the effectiveness of the proposed method, showing significant improvements in image quality and damage localization accuracy.