MUMBAI, India, Sept. 28 -- Intellectual Property India has published a patent application (202641112200 A) filed by Krissal K V; Syed Aejaz Armed A; Nowrin Begum R; and Sam V George on September 18, 2026, for Asrd-U-Mamba: An Adaptive State-Space Deep Learning Framework For Building Footprint Extraction From Aerial Remote Sensing Imagery.
Inventor includes Sam V George.
The application for the patent was published on September 25, 2026, under issue no. 39/2026.
Abstract: Accurate building footprint extraction from aerial remote sensing imagery is essential for urban mapping, spatial planning, infrastructure assessment, and geospatial analysis. However, the complex appearance of rooftops, variations in building size and shape, densely distributed structures, irregular boundaries, and challenging background regions make precise semantic segmentation difficult. This research investigates Mamba-based deep learning architectures for building footprint extraction using the SVAMITVA aerial imagery dataset. A progressive experimental framework is developed by evaluating Original Mamba, U-Mamba, VMamba, 2D U-Mamba, 2D U-Mamba with Deformable Routing (DR), 2D U-Mamba with Deformable Cross Attention (DCA), and the proposed ASRD-U-Mamba architecture. The study begins with established Mamba-based segmentation architectures and progressively introduces architectural modifications to improve spatial feature representation and segmentation of complex building structures. The proposed ASRD-U-Mamba integrates the investigated feature-processing components into a unified architecture designed to capture both long-range contextual information and spatially detailed building features. All models are trained and evaluated under a consistent experimental setting using quantitative metrics including Intersection over Union (IoU), Dice Score, Accuracy, Precision, Recall, and F I - score. In addition to numerical evaluation, qualitative comparisons are performed using identical test images, including visual analysis of segmentation outputs and TP, FP, FN, and TN regions. An ablation-based comparison across the seven architectures is used to determine the contribution of each architectural modification and identify the most effective configuration. The resulting framework provides a systematic approach for improving Mamba-based semantic segmentation for accurate and robust building footprint extraction from aerial remote sensing imagery.
Disclaimer: Curated by HT Syndication.
