Research map / Detection and segmentation / Weakly supervised detection and segmentation
Learning without Exact Guidance: Updating Large-scale High-resolution Land Cover Maps from Low-resolution Historical Labels
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Paraformer, a weakly supervised CNN-Transformer hybrid that updates large-scale 1-m land-cover maps using only low-resolution historical labels (30 m or 10 m), with a pseudo-label-assisted training module that masks inconsistent label pixels.
Problem
Producing up-to-date high-resolution land-cover maps over large areas when only low-resolution historical labels are available, causing spatial mismatch between HR images and LR labels.
Approach
A parallel downsampling-free CNN branch and Transformer branch capture local and global context, and a PLAT module trains on the intersection of the CNN branch's primal prediction and the LR label (mask cross-entropy), so no HR labels or human intervention are needed.
Key results
- On the Chesapeake Bay dataset (30x resolution gap), Paraformer reaches 64.65% average mIoU versus 63.08% for L2HNet, the strongest baseline.
- On the Poland dataset with a 120x gap (GLC FCS30 30 m labels), Paraformer achieves 49.39% mIoU, 5.77 points above L2HNet.
- Ablation: removing PLAT drops average mIoU to 62.81%, and the sole CNN branch gives 60.15% while the sole Transformer branch gives 56.49%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Paraformer instead of L2HNet | large-scale HR land-cover mapping · Chesapeake Bay, six-state average | mIoU (%): 63.08 -> 64.65 | Helped |
| use Paraformer instead of L2HNet | large-scale HR land-cover mapping · Poland, FROM GLC10 | mIoU (%): 6.42% | Helped |
| use Paraformer instead of L2HNet | large-scale HR land-cover mapping · Poland, ESA GLC10 | mIoU (%): 3.06% | Helped |
| use Paraformer instead of L2HNet | large-scale HR land-cover mapping · Poland, Esri GLC10 | mIoU (%): 4.29% | Helped |
| use Paraformer instead of L2HNet | large-scale HR land-cover mapping · Poland, GLC FCS30 | mIoU (%): 5.77% | Helped |
| add the Transformer branch and PLAT to the CNN branch (ablation) | Paraformer · Chesapeake Bay, six-state average | mIoU (%): 60.15 -> 64.65 | Helped |
| add the CNN branch and PLAT to the Transformer branch (ablation) | Paraformer · Chesapeake Bay, six-state average | mIoU (%): 56.49 -> 64.65 | Helped |
| add PLAT to the hybrid CNN-Transformer structure (ablation) | Paraformer · Chesapeake Bay, six-state average | mIoU (%): 62.81 -> 64.65 | Helped |
Surprising findings
- The Transformer-only branch scored lowest in ablation, falling behind even a sole CNN branch, despite its global modeling.
- Paraformer is not best in every state: L2HNet beats it in Virginia and West Virginia on the Chesapeake Bay dataset.
Limitations
- Mask-based training discards inconsistent label pixels, so supervision is limited to pixels where CNN prediction and LR label agree.
- Full model is large (109.4M params, 141.3G FLOPs) relative to the CNN-only variant.
- Evaluated only on two datasets, both with four unified base classes for mIoU.
Experimental setup
Datasets: Chesapeake Bay dataset (NAIP 1 m images, NLCD 30 m labels) Poland dataset (LandCover.ai images, FROM GLC10, ESA GLC10, ESRI GLC10, GLC FCS30 labels) OpenEarthMap ground truths
Benchmarks: Chesapeake Bay dataset Poland dataset
Metrics: mIoU
Baselines: L2HNet TransUNet ConViT CoAtNet MobileViT EfficientViT UNetFormer DC-Swin UNet HRNet LinkNet SkipFCN
Builds on: L2HNet TransUNet UNetFormer
Closest papers by meaning
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.85)
- GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.85)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.84)
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.84)
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.84)
- Spatial-Logic-Aware Weakly Supervised Learning for Flood Mapping on Earth Imagery (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.