<p>Urban scene segmentation of very high-resolution (VHR) remote sensing imagery is a key technology that supports applications such as environmental monitoring, smart city planning, and land use monitoring. However, the complex geometric structures, fine texture features, and diverse land cover objects in VHR images pose challenges to existing methods, including significant background clutter, boundary fragmentation, and insufficient segmentation accuracy of small-scale objects. Existing Mamba-based segmentation networks, such as VM-UNet and Swin-UMamba, rely on purely symmetric SSM encoder–decoder designs without incorporating attention-driven feature recalibration, and their scanning strategies remain limited in capturing omni-directional spatial dependencies, leading to suboptimal performance in cluttered urban scenes. To address the above issues, this paper proposes UMENet, a lightweight VHR semantic segmentation network that introduces a novel Mamba-attention hybrid integration paradigm, where selective state-space modeling is synergistically combined with entropy-aware attention to jointly enhance global context aggregation and channel-wise feature recalibration. First, an entropy-aware channel attention mechanism (EACA) is designed, which replaces conventional global average/max pooling in traditional channel attention with a variance-based information-richness descriptor. Unlike average pooling, which over-smooths spatial details, and max pooling, which can be dominated by isolated strong responses, the variance-based proxy preserves the discriminative spatial distribution characteristics of feature channels, including building edges and road textures, enabling dynamic enhancement of key semantic representations across encoder layers and effectively alleviating the problem of small object omission. Second, a parallel Mamba fusion module (PMF) is proposed, which combines forward, boustrophedon, and circular multi-directional scanning strategies to break through the direction-sensitive limitations of traditional state-space models and improve global context modeling capabilities. Finally, an edge-enhanced deformable dual-head (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {E}^{2}\hbox {D}^{2}\hbox {H}\)</EquationSource> </InlineEquation>) is introduced, integrating deformable convolution with a boundary enhancement mechanism to optimize boundary classification accuracy in cluttered scenes. The network employs a MobileNetV4 efficient encoder and is designed with two versions based on hardware constraints: UMENet-T (with 1.53M parameters) focuses on the efficiency of edge-side deployment, while UMENet-M is an accuracy-oriented high-capacity version. Experiments on the ISPRS Vaihingen and Potsdam datasets show that UMENet-M achieves overall accuracy (OA) of 93.3% and 92.0%, significantly outperforming existing methods. Ablation experiments on both datasets show consistent module contributions, with EACA and PMF providing 0.9% and 1.2% mIoU gains, respectively, and module-substitution experiments further compare EACA with SENet/CBAM/GAP and PMF with Vanilla Mamba. This study provides a feasible solution for high-precision remote sensing segmentation in resource-constrained scenarios.</p>