Spatially Adaptive Multiscale Convolutions With Direction-Aware Hybrid Long-Range Modeling for Remote Sensing Image Superresolution

Abstract

Although self-attention has advanced remote sensing image superresolution (RSISR), its high computational cost and large-scale variations in remote sensing imagery hinder efficient deployment on edge devices. Given this, we propose a lightweight RSISR network, called the direction-aware residual network (DireNet), which synergizes spatially adaptive multiscale convolutions with direction-aware hybrid long-range modeling. DireNet consists of a stack of spatially adaptive feature modulation blocks (SAFMBs), followed by a dual-path hybrid kernel attention module (DPHKAM). Each SAFMB comprises two key components: 1) a grouped adaptive selection attention module (GASAM) that employs multiscale large-kernel dilated convolutions and adaptively selects kernels to dynamically generate multigranularity attention maps for joint local–global modeling and 2) a spatial detail enhancement module (SDEM) that integrates prenormalized residual learning with a lightweight inverted bottleneck architecture, employing depthwise separable convolutions to recover high-frequency details and thereby enhance reconstruction fidelity. The subsequent DPHKAM further enhances global context integration by combining large-kernel attention with horizontal and vertical stripe convolutions, capturing both isotropic and direction-aware long-range dependencies. Experiments show that DireNet achieves comparable reconstruction quality while significantly reducing computational complexity, providing great potential for subsequent edge deployment.