Skip to main content
GIS-Schools GIS-Schools
Change Detection ยท GeoAI

ChangeFormer V6 - LEVIR-CD Change Detection

Compare two dates through a multiscale Siamese Transformer. This page selects the original LEVIR-trained V6 checkpoint and explains its binary building-change output.

Change Detection task illustration
Family: Siamese hierarchical TransformerArchitecture: Multiscale Siamese encoders + difference modules + MLP decoderVersion: ChangeFormerV6 / LEVIR / release v0.1.0License: MIT LICENSE file; README restricts research/non-commercial use - clarify with authors
ChangeFormer V6 - LEVIR-CD Change Detection

Exact checkpoint

Selected official release: v0.1.0, LEVIR ChangeFormerV6 archive with embed_dim_256 in its filename. It is distinct from the DSIFN archive and from an initialization-only pretrained encoder. The release archive is 985,610,406 bytes.

How it works

Shared hierarchical encoders extract several feature resolutions from both dates. Difference modules compare corresponding features. A lightweight decoder combines the scales into a dense change prediction.

Typical input and output

Input is a co-registered RGB pair. The published training and evaluation examples use 256-pixel chips and embed_dim 256 for this checkpoint. Output is binary change segmentation, not a multi-class transition matrix.

Strengths

Multiscale features connect local boundary cues with broader spatial context. A Siamese design provides a practical architecture for controlled two-date comparisons.

GIS and remote-sensing use

Use the result as a candidate building-change layer. Keep unchanged areas for false-positive assessment, inspect changes near tile borders, and carry the common CRS and pixel transform into the final raster. Date-pair provenance is essential for interpreting the output.

Limitations

Registration errors, new shadows and sensor differences may produce false changes. The LEVIR building-change checkpoint does not establish general damage or vegetation-change capability. The original project needs explicit modifications for multi-class change detection.

Training and fine-tuning

Prepare aligned date pairs and binary masks. Use the V6 configuration and checkpoint-compatible embedding size. Review multi-scale training/inference settings and label semantics before adapting the training script. Split by place before tiling and compare against a simple image-difference baseline to expose registration problems.

Framework and hardware

Original PyTorch ChangeFormer implementation. Its scripts assume a configured GPU and research environment. Follow the repository dependency guidance; do not infer a fixed GPU-memory requirement from one published batch size.

Files, license and download

External official release only. The archive exceeds the 100 MiB resource limit. There is also conflicting license wording: LICENSE says MIT while the README limits code to research/non-commercial use. No unrestricted redistribution or commercial permission is asserted; ask the authors to clarify.

Before using a result

  1. Keep imagery rights, acquisition dates and preprocessing with each experiment.
  2. Separate locations before tiling to avoid train/test leakage.
  3. Inspect failures and uncertain outputs alongside the original imagery.
  4. Preserve coordinate transforms and verify units before exporting a GIS layer.

Validation status: Source provenance and website delivery are checked. GISSchools has not run model inference, training or an accuracy benchmark for this record. Feature images and gallery diagrams are explanatory illustrations, not predictions.

Official references

Examples and images

Downloads and resources

External resourceOfficial Model WeightsProvider: W. G. C. Bandara and Vishal M. PatelVersion: ChangeFormerV6 / LEVIR / release v0.1.0License: MIT LICENSE file; README restricts research/non-commercial use - clarify with authors

External official release only. The archive exceeds the 100 MiB resource limit. There is also conflicting license wording: LICENSE says MIT while the README limits code to research/non-commercial use. No unrestricted redistribution or commercial permission is asserted; ask the authors to clarify.

Open resource ↗
External resourceOfficial RepositoryProvider: W. G. C. Bandara and Vishal M. Patel
Open resource ↗
External resourceDocumentationProvider: W. G. C. Bandara and Vishal M. Patel
Open resource ↗
External resourceResearch PaperProvider: W. G. C. Bandara and Vishal M. Patel
Open resource ↗
External resourceLicense and Usage TermsProvider: W. G. C. Bandara and Vishal M. PatelLicense: MIT LICENSE file; README restricts research/non-commercial use - clarify with authors
Open resource ↗
Discussion

Comments

0

No comments yet. Start the discussion.

Join the discussion

Leave a reply

Your email address will not be published. Required fields are marked.