MDFNet: a lightweight multi-stream network tailored for real-time GIS and disaster monitoring applications

Abstract

Very high-resolution remote sensing image scene classification (RSSC) is a key enabling technology for geographic information systems (GIS) in disaster prevention, emergency response, and recovery. Current methods often suffer from excessive complexity, restricting their capability to meet the stringent computational efficiency requirements of time-sensitive edge deployment. To address these structural bottlenecks, this paper presents MDFNet, a lightweight multi-stream dual-stage fusion network based on transfer learning. The model uses parallel low-, middle-, and high-level streams to extract hierarchical features, adopts a dual attention mechanism to enhance discriminative regions, and employs multi-source feature aggregation to reduce redundancy. Extensive experiments on AID, UC Merced, and NWPU-RESISC45 datasets demonstrate that MDFNet achieves competitive or superior accuracy with significantly fewer parameters and lower computational costs. This parameter-efficient design provides a foundational accelerated architecture highly tailored for prospective real-time GIS inference and emergency mapping.