GMBFormer: An NDVI-Guided Global Memory Bank Transformer for Urban Green-Space Extraction from Ultra-High-Resolution Imagery

2026-06-04 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Environmental Science & Earth Systems · Depth: Expert, quick

Summary

GMBFormer is a SegFormer-based framework designed for urban green-space extraction from ultra-high-resolution (UHR) imagery, addressing limitations of patch-by-patch processing and direct Normalized Difference Vegetation Index (NDVI) injection. It replaces adjacency-driven feature propagation with selective, similarity-driven prototype retrieval. The framework decouples NDVI as a physics-informed gate, admitting high-confidence vegetation descriptors into a compact global memory bank via momentum updates, while only RGB channels enter the backbone. During training and inference, the current patch queries stored prototypes through memory-mediated cross-attention. GMBFormer achieved mean intersection over union (mIoU)/mean Dice (mDice) scores of 89.25%/94.31%, 92.17%/95.92%, and 83.72%/90.86% on a Chengdu UHR dataset and two ISPRS Potsdam settings, outperforming the SegFormer-B4 baseline.

Key takeaway

For Computer Vision Engineers developing semantic segmentation models for ultra-high-resolution imagery, particularly in remote sensing, you should consider GMBFormer's approach. Its decoupled NDVI gating and global memory bank effectively overcome the limitations of traditional patch-by-patch processing, improving semantic reuse and overall accuracy. This method offers a robust way to integrate physical indices without blurring their role with visual features, potentially enhancing your model's performance on similar complex datasets.

Key insights

GMBFormer improves urban green-space extraction by decoupling NDVI and using a global memory bank for semantic reuse.

Principles

• Patch-by-patch processing limits semantic reuse among similar patterns.
• Direct NDVI injection blurs roles of visual appearance and physical confidence.

Method

RGB channels process visual appearance; NDVI acts as a gate for high-confidence vegetation descriptors into a momentum-updated global memory bank; patches query prototypes via cross-attention.

In practice

• Implement a global memory bank for cross-patch semantic reuse.
• Decouple physical indices from visual backbones in segmentation.

Topics

• Urban Green-Space Extraction
• Semantic Segmentation
• Transformers
• NDVI
• Memory Networks
• Ultra-High-Resolution Imagery

Best for: AI Scientist, Computer Vision Engineer, Research Scientist

Related on AIssential

• From Semantics to Pixels: Coarse-to-Fine Masked Autoencoders for Hierarchical Visual Understanding — C2FMAE learns hierarchical visual representations by integrating coarse-to-fine masking and cascaded decoding.
• MapSR: Prompt-Driven Land Cover Map Super-Resolution via Vision Foundation Models — MapSR uses prompt-driven vision foundation models for efficient, training-free land-cover super-resolution with minimal annotation.
• NG-GS: NeRF-Guided 3D Gaussian Splatting Segmentation — NG-GS enhances 3D Gaussian Splatting segmentation by integrating NeRF-guided continuous feature fields to resolve boundary artifacts.
• LALE: Lightweight-Transformer Architecture for Land-Cover Estimation — LALE efficiently segments remote sensing imagery by combining ConvMixers for local detail and transformers for global context in a bifurcated encoder.
• HQF-Net: A Hybrid Quantum-Classical Multi-Scale Fusion Network for Remote Sensing Image Segmentation — Hybrid quantum-classical networks can significantly improve remote sensing semantic segmentation by fusing multi-scale features and quantum-enhanced refinement.
• Colour Extraction Pipeline for Odonates using Computer Vision — A computer vision pipeline segments Odonates and extracts color data to enable large-scale ecological correlation analysis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.