LiteLKG-Net: A Lightweight Network With Multiscale Large Kernel Attention and Gated Spatial Feature Fusion for Remote Sensing Change Detection

With the advancement of Earth observation technology, remote sensing change detection (RSCD) has become an effective means to monitor dynamic changes in surface objects, widely applied in urban planning, environmental disaster monitoring, and other fields. Although Transformer-based methods exhibit strong global modeling capabilities, they often suffer from high computational complexity and excessive parameter counts, while traditional CNN-based methods are constrained by limited convolutional kernel sizes, leading to the neglect of global contextual information. These limitations hinder their applicability in resource-constrained and complex scenarios. To address these challenges, this paper proposes a lightweight network (LiteLKG-Net), which integrates multi-scale large kernel attention and a gating mechanism to maintain high detection performance while significantly reducing computational costs. First, we design a Multi-scale Attention Module (MAM) composed of a Multi-scale Large Kernel Attention (MLKA) and a Gated Spatial Attention Unit (GSAU). The MLKA aggregates global and local information to capture rich change features at varying granularities, while the GSAU combines the gating mechanism with spatial attention to fully exploit contextual spatial information with reduced computational overhead. Second, we construct a Channel Interaction Module (CIM) to enhance cross-temporal feature interaction, thereby improving spatiotemporal semantic perception. A Feature Fusion Module (FFM) is further introduced to aggregate multi-scale features, enhancing the model’s capability to discern fine-grained details. The decoder incorporates a multi-level residual fusion strategy and embeds a Gated Mechanism Module (GMM) to adaptively filter critical features, significantly enhancing model accuracy and robustness. Finally, we adopt MobileNetV2 as the Siamese encoder to efficiently extract features while minimizing computational complexity. Comprehensive experiments are conducted on three benchmark datasets. Both visual and quantitative results demonstrate that our proposed LiteLKG-Net outperforms ten other state-of-the-art networks. Additionally, ablation studies validate the effectiveness of each proposed module.