LDSINet: Lightweight Dual‐Stream Integration Network for Salient Object Detection in Optical Remote Sensing Images

ABSTRACT Salient object detection in remote sensing images (RSI‐SOD) plays a crucial role in Earth observation and intelligent interpretation. However, existing methods are often computationally intensive and rely on large models, making it challenging to balance accuracy and efficiency. In this paper, we propose LDSINet, which achieves efficient feature modeling through three key optimizations: (1) Introducing a novel Mamba‐based branch to replace the traditional PVT branch, retaining global modeling capability while reducing computational overhead; (2) Designing an efficient pyramid pooling module to enable low‐cost multi‐scale context fusion; (3) Developing an efficient nested cross‐layer decoder (ENCD) for compact multilevel feature integration. Experimental results demonstrate that LDSINet, with 63.85M parameters and 8.66G FLOPs, significantly outperforms existing efficient networks in detection accuracy, while achieving comparable precision to the high‐accuracy DSINet baseline with substantially reduced parameter count and computational cost. Moreover, LDSINet provides accelerated inference on inputs. This approach offers an efficient solution for high‐precision, resource‐aware RSI‐SOD.