Satellite damage-assessment pipeline compresses reference imagery by up to 64 times

Researchers report a satellite building-damage assessment pipeline that compresses pre-disaster reference images by up to a factor of 64 and returns building locations and damage classes instead of full scenes. The arXiv preprint says strongly compressed references preserve most detection performance on the xBD dataset. The system compares compressed reference features with new imagery and corrects alignment errors before combining them. Researchers ported it to a Xilinx Versal VCK190 and an NVIDIA Jetson AGX Orin. The core detector worked on both, but operators needed for robustness to large alignment shifts worked only on the Jetson GPU.