Scalable Urban Semantic Segmentation and Building Footprint Extraction via Linear-Time Feature Fusion

Accurate urban infrastructure mapping and building footprint extraction from high-resolution aerial tiles require macro-geographical context to be captured while precise boundary definitions are maintained.However, traditional Transformer-based encoders suffer from quadratic computational and memory complexity, O(N²), which severely limits their scalability on extensive metropolitan mosaics.In this study, we propose the Edge-Enhanced Context-Aware Mamba Network (EECANet-Mamba), a fully linear-time context-aware network designed for scalable, high-throughput urban semantic segmentation.At its core, the Mamba-Guided Selective Scan Fusion (M-SSF) module aligns hierarchical backbone features and models global spatial dependencies in linear time, O(N), using multi-directional selective state-space scanning.To prevent localized urban borders from being smeared by the sequential low-pass filtering characteristic of Mamba scans, an asymmetric gating mechanism learned directly from the shallow, highresolution feature map is introduced.This gate routes a boundary-emphasized representation to an Edge-Aware Refinement Module (EARM) under auxiliary edge supervision, while direct context is preserved via a High-Resolution Preservation Pathway (HRPP).Strict linear scalability is confirmed by rigorous theoretical analysis and physical hardware profiling on an NVIDIA A100 GPU, demonstrating a mere 32 ms wall-clock latency and a 600 footprint at 1,024 × 1,024.Benchmark evaluations on GID-15, LoveDA, and ISPRS Vaihingen achieve 69.7%, 57.1%, and 83.7% mean Intersection over Union (mIoU), respectively.On the complex urban layouts of Vaihingen, the network yields an exceptional 94.9% F1-score on buildings and 84.1% on vehicles while maintaining a high throughput of 80 frames per second (FPS), establishing an optimal accuracy-efficiency frontier for scalable, regional Earth observation pipelines.