Lightweight AI Framework Boost SAR Building Segmentation Accuracy and Speed

Accurate building extraction from Synthetic Aperture Radar (SAR) imagery is a critical requirement for modern urban mapping, sophisticated disaster assessment, and advanced remote-sensing interpretation. However, the inherent nature of SAR data presents significant challenges to automated analysis. These difficulties stem from several factors, including pervasive speckle noise, heterogeneous scattering effects, the presence of radar shadows, weak textural features, and inherently ambiguous building boundaries. These issues collectively degrade the quality of pixel-level predictions, making robust and reliable automated segmentation extremely difficult. Compounding this problem is the practical constraint that heavy, complex segmentation networks, while powerful, are often too computationally demanding to be deployed in resource-constrained environments, such as edge devices or large-scale real-time monitoring systems. Addressing this gap, researchers have introduced a novel solution: a task-oriented lightweight SAR building segmentation framework.

This innovative framework is derived from the established DeepLabV3+ architecture but is specifically optimized for efficiency and performance in SAR environments. The core of the solution lies in a strategic combination of highly efficient architectural components designed to tackle the specific limitations of radar data while minimizing computational overhead. Central to this design is the utilization of MobileNetV3 as the backbone.

To effectively capture the complex spatial relationships and varying sizes of buildings—a necessity for accurate segmentation—the framework incorporates a depthwise separable atrous spatial pyramid pooling (DS-ASPP) module. The DS-ASPP module is a sophisticated mechanism that retains crucial multi-scale context within the image data. By employing depthwise separable convolutions, the module achieves this multi-scale context at a substantially reduced computational cost compared to traditional, heavier pyramid pooling methods. This allows the model to perceive both fine-grained local details and broad global structures simultaneously, which is vital for distinguishing building shapes from surrounding urban clutter in SAR imagery.

Furthermore, the framework introduces a specialized bidirectional guided fusion module (BGFM). This module is engineered to perform a critical task: combining two distinct types of information. It integrates boundary-sensitive low-level information—which is crucial for defining precise building edges—with deep semantic features. By guiding the fusion process bidirectionally, the BGFM ensures that the model leverages both the fine-grained structural cues and the high-level contextual understanding derived from the deeper layers of the network, leading to more accurate and well-defined segmentation masks.

The performance of this proposed model was rigorously tested on the SARBuD1.0 dataset. The results demonstrated a highly favorable accuracy-efficiency trade-off. On SARBuD1.0, the model achieved an overall accuracy (OA) of 94.64%, a mean Intersection over Union (MIoU) of 79.89%, and a Kappa score of 76.45%. Crucially, this high level of performance was achieved using only 11.3 million parameters, indicating a highly optimized model size. In terms of speed, the model operated at 58.12 Frames Per Second (FPS), showcasing its efficiency for real-time applications.

When compared against existing state-of-the-art models, the proposed framework demonstrated superior performance. Across five independent runs on SARBuD1.0, the proposed model achieved a mean MIoU of 79.86 ± 0.17%. This result was statistically significantly higher than the 78.52 ± 0.20% achieved by HSB-Net, with a significant paired difference confirmed at p < 0.001. When tested on the HR-SARBuD dataset, the model achieved a MIoU of 79.75%. The five-run mean for this dataset was 79.77 ± 0.15%, which was compared against FPANet’s mean of 79.68 ± 0.16%, though this difference was not statistically significant (p = 0.214).

The computational efficiency of the framework is perhaps one of its most compelling features. Under a batch-1 complexity protocol, the proposed network requires only 8.9 GFLOPs and processes images in 17.21 milliseconds. This stands in stark contrast to the DeepLabV3+ architecture, which requires 42.7 GFLOPs and takes 50.89 milliseconds per image. These metrics clearly support the claim of a favorable accuracy-efficiency trade-off, making the model viable for deployment on edge devices and in computational-constrained settings where heavy segmentation networks are typically prohibitive.

The research team explicitly discussed the limitations of the model, particularly concerning cross-dataset performance and potential edge-device constraints, demonstrating a thorough and responsible approach to the deployment of the technology. By successfully integrating efficient backbones like MobileNetV3 with specialized modules such as DS-ASPP and BGFM, the researchers have provided a powerful tool for remote sensing applications that can operate effectively where computational resources are limited. This advancement moves the field closer to real-time, large-scale urban monitoring using SAR data.

Subject of Research: A lightweight deep learning framework for accurate building segmentation in Synthetic Aperture Radar (SAR) imagery.

Article Title: Lightweight SAR building segmentation via enhanced multi-scale context modeling and bidirectional guided feature fusion

Article References: Wang, Q., Long, Z., Chen, Z., Liang, X., & Xu, X. (2026). Lightweight SAR building segmentation via enhanced multi-scale context modeling and bidirectional guided feature fusion. Scientific Reports. https://doi.org/10.1038/s41598-026-76312-3

Image Credits: AI Generated

DOI: 10.1038/s41598-026-76312-3

Keywords: SAR, Deep Learning, Computer Vision, Building Segmentation, MobileNetV3, Remote Sensing, AI Efficiency, Lightweight, building, segmentation, enhanced, multi-scale