Research map / Vision-language and multimodal / Multimodal LLMs

Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 1 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 Landsat30-AU, a 30 m multi-satellite (Landsat 5/7/8/9) Australian vision-language dataset with 196k captions and 17.7k human-verified VQA items, and benchmarks VLMs, showing LoRA fine-tuning of Qwen2.5-VL-7B greatly improves captioning and VQA.

Problem

Existing remote-sensing vision-language datasets focus on sub-meter commercial imagery from few satellites and short time spans, leaving long-term, low-resolution, multi-satellite Landsat archives without image-text supervision.

Approach

A bootstrapped pipeline that uses noisy but spatially/temporally aligned metadata (DEA land cover, OSM tags) plus generic VLM drafting, fine-tuned region/caption/review modules, iterative caption refinement, and human verification to create resolution-aware Landsat captions and VQA.

Key results

• EarthDial scores 0.07 SPIDEr on captioning and 0.48 overall VQA accuracy, while Qwen-ft improves SPIDEr from 0.11 to 0.31.
• Qwen-ft reaches 0.871 overall VQA accuracy versus 0.743 for base Qwen2.5-VL-7B.
• The fine-tuned GPT-4.1 caption generator raised SPIDEr from 0.44 to 0.52 on the fine-tune set.

Surprising findings

• Numerosity (NUM) is a universal bottleneck across all evaluated VLMs, with the best overall model at about 0.66 accuracy on that category after fine-tuning.
• RS-LLaVA, a remote-sensing VLM, achieves the lowest SRI and USR scores among models, suggesting domain-trained specialized models may not transfer to Landsat imagery.

Limitations

• Human verification covers only a subset of the data; large-scale captions rely on automated VLM refinement and review.
• Captions are derived from generic VLM drafts with OSM and land-cover metadata, so hallucinations and metadata errors may remain.
• Benchmark covers Australia only and uses 256x256 RGB tiles at 30 m.

Experimental setup

Datasets: Digital Earth Australia ARD Landsat imagery DEA Land Cover OpenStreetMap tags

Benchmarks: Landsat30-AU-Cap Landsat30-AU-VQA

Metrics: BLEU-4 SPIDEr BERTScore-F1 1-CHAIR-s 1-CHAIR-i avg caption length per-category VQA accuracy overall VQA accuracy

Baselines: EarthDial RS-LLaVA MiMo-VL GLM-4.1V Gemma 3 Llama-3.2 LLaVA-OneVision

Builds on: LHRS-Align VRSBench ChatEarthNet LoRA

Closest papers by meaning

• GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.90)
• SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling (CVPR 2025 · similarity 0.90)
• VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding (NeurIPS 2024 · similarity 0.90)
• SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing (AAAI 2024 · similarity 0.90)
• EarthDial: Turning Multi-sensory Earth Observations to Interactive Dialogues (CVPR 2025 · similarity 0.89)
• TEOChat: A Large Vision-Language Assistant for Temporal Earth Observation Data (ICLR 2025 · similarity 0.89)