GeoMeld: 2.5 Million Multi-Modal Remote Sensing Samples

GeoMeld is a large-scale multi-modal remote sensing dataset introduced in a CVPRW 2026 paper. It contains approximately 2.5 million spatially aligned samples spanning heterogeneous sensing modalities and resolutions, paired with semantically grounded captions. The dataset was created by author vimageiitb and last updated on 2026-06-04.

Use Cases

• Training multi-modal foundation models based on spatially aligned remote sensing samples.
• Semantically grounded image captioning based on agentically generated text descriptions.
• Cross-modal retrieval tasks based on paired image-text data from heterogeneous sensing modalities.
• Benchmarking models for earth observation based on the dataset's scale and multi-modal nature.

Strengths

• Approximately 2.5 million samples provides a substantial scale for training.
• Spatial alignment across samples suggests consistent georeferencing.
• Multi-modal pairing of remote sensing data with semantically grounded captions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

vimageiitb

Collection Method

Samples paired with captions generated through an agentic pipeline.

Freshness

Last updated 2026-06-04 16:00:39; freshness should be verified.