Landsat30-AU: Multisensor Landsat Imagery in Australia

• Landsat30-AU is a large-scale, resolution-aware resource pairing 30 m Landsat imagery with human-verified captions and VQA for long-term analysis.
• It leverages a structured human-in-the-loop pipeline to address sensor diversity, temporal continuity, and regional adaptation over Australia.
• In burned-area mapping, it adapts the GABAM 2015 product using automated detection and Random Forest classification on Google Earth Engine.

Landsat30-AU denotes a large-scale, resolution-aware vision–language resource built for long-term, multi-satellite Landsat imagery over Australia, and, in a separate usage, the term has also referred to applying the 30 m Landsat-based Global Annual Burned Area Mapping product and methods to Australian conditions. In the vision–language sense introduced by Sai Ma, Zhuang Li, and John A. Taylor, the dataset addresses a gap in remote sensing VLM resources by pairing 30 m Landsat imagery with human-reviewed captions and VQA designed explicitly for Landsat-style reasoning across more than 36 years. In the burned-area sense, the term has been used for Australian use of GABAM 2015, a global annual burned area map derived from Landsat imagery and implemented entirely on Google Earth Engine (Ma et al., 5 Aug 2025, Long et al., 2018).

1. Definition, scope, and disambiguation

Landsat30-AU, as defined in the 2025 work, is a large-scale, resolution-aware vision–language corpus purpose-built for long-term, multi-satellite Landsat imagery over Australia. Its stated motivation is that most prior datasets emphasize short-term, sub-meter commercial or Sentinel-2 sources, whereas Landsat offers an affordable, bias-robust global archive at 30 m ground-sample distance with decades of continuity across multiple missions. The dataset therefore targets sensor diversity, mission-dependent color shifts, coarse spatial detail, and multi-decadal temporal breadth, all of which are central to Landsat-style interpretation (Ma et al., 5 Aug 2025).

This scope differs from high-resolution remote-sensing caption corpora in a fundamental way. At 30 m GSD, fine-grained objects typical of sub-meter datasets are absent; models must instead reason about meso-scale texture, spatial layout, and sensor-specific colorimetry. The dataset is explicitly designed around that regime, with captions and VQA curated to reference objects and relations visible at 30 m and with human verification intended to prevent ungrounded or temporally mismatched text. Relative to prior resources, the paper positions Landsat30-AU against UCM-Captions, Sydney-Captions, RSICD, NWPU-Captions, RS5M/GeoRSCLIP, Git-10M, RS-LLaVA, ChatEarthNet, SkyScript, GAIA, EarthDial, and SSL4EO-L, emphasizing that Landsat content, when present, is often limited in satellite diversity or geolocation metadata (Ma et al., 5 Aug 2025).

A separate, narrower usage associates Landsat30-AU with Australian exploitation of the 30 m Landsat-based GABAM 2015 burned-area product. In that usage, the emphasis is not image–text alignment but automated annual burned-area delineation from Landsat-8 time series in GEE, including explicit separation of new burns from legacy scars and application guidance for Australian rangeland, savanna, and cropland contexts (Long et al., 2018).

2. Dataset composition and source data

The vision–language dataset is Australia-wide, georeferenced, and temporally extensive. It spans 1988–2024, uses Landsat 5, 7, 8, and 9 imagery, and operates at 30 m GSD with 256 × 256 pixel tiles. The imagery is sourced from Digital Earth Australia Analysis-Ready Data, which is atmospherically and geometrically corrected. Captions use true-color RGB from nbart_red, nbart_green, and nbart_blue; panchromatic band pansharpening to 15 m was used during human verification only and is not distributed as part of the dataset. Per-image metadata retain geohash, acquisition date, and satellite information (Ma et al., 5 Aug 2025).

The data preparation procedure is highly structured. Australia is partitioned into approximately 2,500 static Areas of Interest, each approximately $7{,}680 \times 7{,}680\,\mathrm{m}^2$. Temporal sampling targets one image per AOI per year, with quarter cycling to diversify seasons, and the first image passing quality filters is selected. Initial filtering retained patches marked by ARD metadata as at least 99.5% cloud-free, after which Qwen2.5-VL-7B classified each patch as “cloudy” or “clear” to remove residual clouds and artifacts. The final dataset contains 196,262 high-quality RGB tiles (Ma et al., 5 Aug 2025).

Two auxiliary semantic sources are central to the curation process. DEA Land Cover v2.0.0 is used to extract dominant land cover for six fixed regions within each tile—top-left, top-right, bottom-left, bottom-right, center, and entire tile—with classes such as Cultivated Terrestrial Vegetation, Natural Terrestrial Vegetation, Natural Aquatic Vegetation, Artificial Surfaces, Natural Bare Surfaces, Water, and No Data. OSM tags are mapped into 25 coarser, Landsat-visible categories to mitigate granularity and terminology inconsistencies. The paper does not specify radiometric normalization beyond using ARD nbart bands, and it does not explicitly state on-disk image formats or CRS, although georeferencing and timestamps are retained per image (Ma et al., 5 Aug 2025).

Component	Count	Notes
Landsat30-AU-Cap	196,262	Final “Reviewed” image–caption pairs
Landsat30-AU-VQA	17,725	Human-verified multiple-choice VQA samples across eight domains

3. Annotation pipeline and task taxonomy

The annotation workflow is a three-stage, human-in-the-loop pipeline. The first stage performs imagery and metadata preparation through DEA ARD sampling, OSM mapping, and DEA Land Cover extraction. The second stage adapts generic VLMs to Landsat using small, human-verified sets. The third stage generates captions and VQA through iterative refinement and review. This architecture is intended to scale annotation while preserving grounding at 30 m GSD (Ma et al., 5 Aug 2025).

The adaptation stage comprises three supervised subproblems. Region classification uses DEA Land Cover labels for the six tile zones; 2,722 tile-region sets were manually validated, and GPT-4o was fine-tuned, achieving Subset Accuracy 0.278 and Jaccard 0.630. Caption generation uses 1,005 image–caption pairs curated with explicit 30 m visibility and date alignment; GPT-4.1 fine-tuning improved SPIDEr from 0.44 to 0.52, 1-CHAIR-s from 0.44 to 0.47, and average length from 149 to 161 tokens. Caption review uses 9,440 sentence-level keep/delete labels, with Qwen2.5-VL-7B fine-tuned to prune hallucinated or temporally inconsistent text (Ma et al., 5 Aug 2025).

The final caption pipeline is multi-stage. “Initial” captions are produced by fine-tuned GPT-4.1, “Extra” captions are expanded by Qwen2.5-VL-7B to add missing objects and relations, and “Reviewed” captions are pruned by a reviewer model. The “Reviewed” variant is used in Landsat30-AU-Cap and achieves SPIDEr 0.517 with 1-CHAIR-i 0.853. Human verification guidelines require that only sentences visually verifiable at 30 m be kept; when ambiguous, reviewers consult high-resolution map sources such as Google Earth to confirm macro-objects including golf courses, small dams, and solar farms. The paper does not report inter-annotator agreement statistics (Ma et al., 5 Aug 2025).

The VQA component comprises 17,725 validated multiple-choice questions across eight task domains. These domains are designed to emphasize what is feasible and what is not feasible at 30 m GSD, including crop-season inference, cloud usability, dominant land cover, detectability of thin structures, macro-object presence, numerosity, spatial relations, and settlement scale. The taxonomy is not merely descriptive; it encodes Landsat-specific perceptual limits and scene abstractions.

Domain	Count	Task focus
APR	2,102	Agro-Phenology Reasoning
COA	2,129	Cloud-Occlusion Assessment
DLC	2,479	Dominant Land Cover
FOD	2,000	Fine-Object Detectability
MOP	2,418	Macro-Object Presence
NUM	2,244	Numerosity
SRI	2,419	Spatial-Relation Inference
USR	1,934	Urban-Scale Recognition

4. Benchmarking, metrics, and model adaptation

The benchmark includes captioning and multiple-choice VQA. Captioning is evaluated on a test set of 1,005 human-verified image–caption pairs with BLEU-4, SPIDEr, BERTScore-F1, 1-CHAIR-s/i, and average caption length. VQA uses exact-match accuracy on a 15% test split of Landsat30-AU-VQA. The paper defines SPIDEr as

$$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$

and VQA accuracy as

$$\mathrm{Acc} = \frac{\#\text{ correct}}{\#\text{ total}}.$$

CHAIR is reported as $1-\mathrm{CHAIR}$, so higher values indicate fewer hallucinations (Ma et al., 5 Aug 2025).

Eight VLMs are evaluated: EarthDial, RS-LLaVA, MiMo-VL, GLM-4.1V, Qwen2.5-VL-7B, LLaVA-OneVision, Llama-3.2, and Gemma 3, plus fine-tuned Qwen-ft and Llama-ft. The central empirical result is that off-the-shelf models struggle with Landsat imagery. EarthDial achieves captioning SPIDEr 0.0726 and overall VQA accuracy 0.4829, with particularly low APR 0.2349 and COA 0.1034. Qwen2.5-VL-7B off the shelf reaches SPIDEr 0.1114 and VQA accuracy 0.7428. After lightweight fine-tuning on Landsat30-AU, Qwen-ft reaches SPIDEr 0.3054 and VQA accuracy 0.8710; Llama-ft reaches SPIDEr 0.2767 and VQA overall 0.7315. This suggests that resolution-aware, sensor-diverse supervision materially improves 30 m GSD reasoning (Ma et al., 5 Aug 2025).

The fine-tuning recipe is deliberately lightweight. QLoRA uses 4-bit NormalFloat quantization, adapter rank 64 with $\alpha=128$ and dropout 0.05, and inserts adapters into attention projections, the gated-MLP stack, and the cross-modal vision projector. Training uses one epoch, AdamW with learning rate $2\times10^{-4}$, cosine annealing, warm-up 6%, and weight decay $10^{-6}$, on 15% of the task-specific training subsets. In parallel, the caption-review model uses full-parameter fine-tuning of Qwen2.5-VL-7B for 3 epochs with batch size 24, Adam, cosine LR schedule starting at $2\times10^{-5}$, warm-up 6%, weight decay $1\times10^{-6}$, $\beta_2=0.999$, and bfloat16 on eight NVIDIA L4-24G GPUs (Ma et al., 5 Aug 2025).

5. Burned-area mapping usage in Australia

In a distinct line of usage, Landsat30-AU refers to use of the 30 m Landsat-based Global Annual Burned Area Mapping product and pipeline in Australia. The underlying product, GABAM 2015, is a 30 m resolution Global Annual Burned Area Map produced from Landsat imagery via an automated pipeline in Google Earth Engine. Its spatial coverage is global, tiled in 10×10 degree units spanning 180W–180E and 80N–60S, with native projection geographic (Lat/Long, WGS84) at 0.00025 degrees, approximately 30 m. Temporally, the released product covers 2015 only, and “annual” is defined strictly as fires that occurred during the target year; fires from previous years that have not fully recovered are explicitly excluded (Long et al., 2018).

The GEE pipeline has three major stages: model training, per-pixel processing over dense time series, and burned-area shaping. It uses USGS Landsat-8 Surface Reflectance Tier 1 and Tier 2 collections in GEE, with QA masking based on Fmask to remove clouds, cloud shadows, water, snow/ice, and filled or dropped pixels. Training relies on 120 Landsat-8 scenes selected by stratified random sampling across land cover and fire-density strata, yielding 6,735 burned and 6,146 unburned samples globally. Fourteen Landsat features are used at each observation time: six bands and eight spectral indices. Feature importance analysis identified NIR, SWIR2, and the indices NBR2, BAI, MIRBI, and SAVI as most informative. Classification is performed by a Random Forest in GEE with 100 trees in probability mode (Long et al., 2018).

The temporal decision logic is designed to suppress false positives and separate new burns from legacy scars. For each 2015 pixel, the method computes per-observation indices and RF burn probabilities, identifies the date $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$0 of maximum burn probability $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$1, and compares event-time NDVI and NBR to prior-year and two-year extrema. Four globally applied filters are then used: $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$2, $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$3, $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$4, and a temporal rule requiring either $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$5 or $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$6 days. The thresholds are $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$7, $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$8, $$\mathrm{SPIDEr} = \tfrac{1}{2}\big(\mathrm{SPICE} + \mathrm{CIDEr}\big),$$9, and $$\mathrm{Acc} = \frac{\#\text{ correct}}{\#\text{ total}}.$$0 days. MODIS VCF is used to distinguish tree-dominated and herbaceous surfaces; for herbaceous vegetation, only the NDVI filters are applied because grasslands can recover very quickly and can burn repeatedly year after year (Long et al., 2018).

Post-processing sharpens the scar geometry. Pixels passing the filters and having $$\mathrm{Acc} = \frac{\#\text{ correct}}{\#\text{ total}}.$$1 become burn seeds. Eight-connected components smaller than 11 pixels, approximately less than 1 ha, are removed, and iterative region growing adds eight-connected neighbors with burn probability at least 0.5 until convergence. Cross-comparison with Fire_cci v5.0 yields a similar spatial distribution and a global $$\mathrm{Acc} = \frac{\#\text{ correct}}{\#\text{ total}}.$$2. Preliminary global validation reports commission error 13.17%, omission error 30.13%, and overall accuracy 93.92%. In an Australian Landsat-8 validation site, path/row 104/074, commission error is 0.77%, omission error 20.88%, and overall accuracy 90.22%, which the paper associates with good performance in rangeland and savanna contexts typical of northern Australia (Long et al., 2018).

6. Limitations, access, and future directions

The Landsat30-AU vision–language dataset is deliberately resolution-aware, but that design also formalizes its limits. At 30 m GSD, thin and sub-pixel structures are not reliably detectable; the VQA taxonomy includes Fine-Object Detectability precisely to encode “what cannot be seen.” Mission-specific color and radiometric shifts remain a challenge, and robust performance is achieved through multi-sensor exposure and fine-tuning rather than explicit cross-sensor normalization. Cloud coverage and seasonal sampling can under-represent persistently cloudy regions, including tropical wet-season cases. OSM introduces possible temporal and spatial mismatch, mitigated through cross-validation, pruning, and human review but not eliminated. The paper also notes that inter-annotator agreement is not reported, and it states that geospatial data should be used responsibly, avoiding inference of sensitive population attributes (Ma et al., 5 Aug 2025).

Access and licensing are partly explicit and partly deferred to repositories. For the vision–language dataset, code and data are available at https://github.com/papersubmit1/landsat30-au. DEA Land Cover used in curation is CC BY 4.0, but the paper does not explicitly state the overall dataset license and directs readers to the repository for definitive terms. Geolocation and timestamps are retained per image, although sensitive metadata were masked in some VQA generation steps to avoid leakage during prompting. The paper identifies Landsat Next, multispectral text supervision, broader geographies, and temporal reasoning benchmarks as natural extensions (Ma et al., 5 Aug 2025).

For the burned-area usage, the GABAM 2015 product can be freely downloaded at https://vapd.gitlab.io/post/gabam2015/, though users are advised to consult the website for license specifics prior to redistribution. Its stated limitations include Landsat revisit and cloud-related temporal gaps, omission in fast-recovering vegetation, confusion in croplands and dark soils, and the use of a single global threshold set across diverse biomes. In Australian applications, the paper recommends ancillary masks for cropland and related confusing classes, aggregation to ecological or administrative units, local validation in tropical savanna, arid rangeland, and temperate forest biomes, and complementary use of coarse-resolution fire products or national fire-scar datasets. The authors also plan long time-series 30 m GABAMs beyond 2015 and more region-specific algorithms and ancillary data use (Long et al., 2018).