Research map / Vision-language and multimodal / Multimodal LLMs

FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

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.

Builds a SAR-specific VLM (FUSAR-GPT on Qwen2.5-VL-7B) that injects AlphaEarth geospatial embeddings via a Gaussian-weighted token-wise linear modulation module and trains in two decoupled SFT stages.

Problem

General VLMs transfer poorly to SAR imagery due to modality gap, scattering-driven sparsity, and lack of geospatial priors and high-quality SAR text corpora.

Approach

Injects AlphaEarth (AEF) 64-d geospatial embeddings, sampled at georeferenced anchors, into the SAR visual tokens via a TLM module that predicts per-channel scale and shift from Gaussian-interpolated priors, and decouples knowledge injection (SFT stage 1, descriptive captions, LoRA + AEF MLP) from task SFT (stage 2, LoRA only).

Key results

• Target counting accuracy 52.53% for FUSAR-GPT vs 45.45% for the best baseline (Qwen3-VL-4B).
• Target detection F1 at IoU 0.25 rises from 47.1% (Qwen2.5-VL-7B) to 74.8%.
• Ablation: Qwen2.5-VL with SFT2 only reaches 34.85%; adding TLM gives 41.92%; full model with SFT1 reaches 52.53%.

Surprising findings

• Larger general-purpose models do not improve SAR counting: Qwen3-VL-4B (45.45%) outperforms Qwen3-VL-8B (41.41%).
• Proprietary ChatGPT-5.2 (18.18%) and Gemini-3 (30.3%) score below the open baselines on SAR counting.

Limitations

• Task-specific evaluation covers only a ~2k-image subset; the paper does not state a public benchmark release.
• Temporal dynamics are not modeled; AEF is used as a static annual prior.
• Detection is not the method's design focus, with only a preliminary comparison against R3Det.

Experimental setup

Datasets: FUSAR-GEOVL-1M FUSAR-GEOVL (10k AEF-image-text triplets) AlphaEarth features

Benchmarks: Target counting (SAR) Spatial localization (3x3 grid) Target classification (plane/ship fine-grained) Target detection (SAR planes/ships) SAR corpus-level captioning

Metrics: accuracy Acc@100 Acc@50 Top1 classification accuracy precision recall F1 BLEU CIDEr

Baselines: Qwen2-VL Qwen2.5-VL Qwen3-VL LLaVA-1.5 LLaVA-1.6 InternVL-3.5 SARLANG-1M SAR-CLIP GeoChat R3Det

Builds on: Qwen2.5-VL LoRA AlphaEarth Foundations FUSAR-KLIP

Closest papers by meaning

• SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection (NeurIPS 2024 · similarity 0.86)
• Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery (AAAI 2026 · similarity 0.85)
• GeoChat: Grounded Large Vision-Language Model for Remote Sensing (CVPR 2024 · similarity 0.84)
• SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling (CVPR 2025 · similarity 0.84)
• VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding (NeurIPS 2024 · similarity 0.84)
• DisasterM3: A Remote Sensing Vision-Language Dataset for Disaster Damage Assessment and Response (NeurIPS 2025 · similarity 0.84)