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

2025-08-05

AI Summary

Existing vision-language model (VLM) datasets predominantly focus on short-term, high-resolution, single-source satellite imagery, limiting their utility for robust, long-term global land surface monitoring. Method: We introduce Landsat30-AU—the first Australia-focused, multi-temporal vision-language dataset spanning 36 years, integrating observations from four Landsat satellites at 30 m spatial resolution. It comprises two task-specific subsets: image captioning (Landsat30-AU-Cap) and visual question answering (Landsat30-AU-VQA). We propose a novel bootstrapping pipeline that synergistically combines general-purpose VLMs, iterative refinement, and human verification to construct high-quality annotations for long-archive, low-resolution, multi-sensor remote sensing data. Contribution/Results: Benchmark experiments demonstrate that lightweight fine-tuning of Qwen2.5-VL-7B achieves substantial improvements: SPIDEr score increases from 0.11 to 0.31, and VQA accuracy rises from 0.74 to 0.87—validating the dataset’s effectiveness and learnability for long-term Earth observation tasks.

Technology Category

• Computer Vision: Language and Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Vision language models (VLMs) that enable natural language interaction with satellite imagery can democratize Earth observation by accelerating expert workflows, making data accessible to non-specialists, and enabling planet-scale automation. However, existing datasets focus mainly on short-term, high-resolution imagery from a limited number of satellites, overlooking low-resolution, multi-satellite, long-term archives, such as Landsat, that are essential for affordable and bias-robust global monitoring. We address this gap with Landsat30-AU, a large-scale vision-language dataset built from 30-meter resolution imagery collected by four Landsat satellites (5, 7, 8, and 9) over Australia, spanning more than 36 years. The dataset includes two components: Landsat30-AU-Cap, containing 196,262 image-caption pairs, and Landsat30-AU-VQA, comprising 17,725 human-verified visual question answering (VQA) samples across eight remote sensing domains. Both datasets are curated through a bootstrapped pipeline that leverages generic VLMs with iterative refinement and human verification to ensure quality. Our evaluation of eight VLMs on our benchmark reveals that off-the-shelf models struggle to understand satellite imagery. The open-source remote-sensing VLM EarthDial achieves only 0.07 SPIDEr in captioning and a VQA accuracy of 0.48, highlighting the limitations of current approaches. Encouragingly, lightweight fine-tuning of Qwen2.5-VL-7B on Landsat30-AU improves captioning performance from 0.11 to 0.31 SPIDEr and boosts VQA accuracy from extbf{0.74} to 0.87. Code and data are available at https://github.com/papersubmit1/landsat30-au.

Problem

• Lack of vision-language datasets for long-term, low-resolution satellite imagery
• Existing models struggle with satellite imagery understanding and accuracy
• Need for affordable, bias-robust global monitoring solutions

Innovation

• Landsat30-AU dataset with 36-year multi-satellite imagery
• Bootstrapped pipeline for quality image-caption and VQA data
• Fine-tuning Qwen2.5-VL-7B improves captioning and VQA performance