A Vision-Language Framework for Multispectral Scene Representation Using Language-Grounded Features

Aykut Erdem; Enes Karanfil; Erkut Erdem; Nevrez Imamoglu

REVIEW 3 major objections 5 minor 30 references

A Vision-Language Framework for Multispectral Scene Representation Using Language-Grounded Features

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper aims to show that aligning frozen SpectralGPT features with text produces multispectral representations that classify and describe scenes better than raw visual features.

desk verdict A useful LLaVA-style port to multispectral remote sensing with a real new dataset, but the headline claim that language grounding drives the classification gain is not yet secured because a non-language supervised control projector is missing. read the letter →

arxiv 2501.10144 v1 pith:WZQVQ3Y7 submitted 2025-01-17 cs.CV

Enes Karanfil , Nevrez Imamoglu , Erkut Erdem , Aykut Erdem This is my paper · ORCID

classification cs.CV MSC 68T4568T5068U10

keywords multispectralremotesensingvision-languagealignmentlanguage-groundedfeatureslinearprobingsceneclassificationdescriptionSpectralGPTBigEarthNet

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Spectral-LLaVA is a vision-language framework that takes a frozen SpectralGPT encoder, maps its multispectral features through a trainable linear projection into the word-embedding space of LLaMA3, and fine-tunes the language model with LoRA. The paper's central claim is that this language-grounded alignment turns raw multispectral features into semantically richer representations: linear probing on EuroSAT gives higher accuracy for aligned features than for raw SpectralGPT features at every training-split ratio, and the scene-description model scores higher than the RGB-based GeoChat baseline on the revised LLaVA-bench subset. The larger point is that multispectral data can be connected to large language models with only a lightweight projector, and that text alignment itself improves representation quality without retraining the vision backbone. If correct, this offers a cheap route to better remote-sensing classifiers and captioners.

What carries the argument

The load-bearing component is the trainable linear projection layer that maps frozen SpectralGPT visual tokens into the LLaMA3 word-embedding space, trained against text from class labels or captions; the resulting language-grounded features are then used both for linear probing and as visual tokens for the LoRA-fine-tuned decoder. Because the vision encoder never updates, any measured improvement is attributed to the alignment itself rather than extra visual pretraining, which is what lets the paper isolate the effect of language grounding.

What would settle it

A decisive check is to retrain the same linear projector with caption texts randomly permuted across images and repeat the EuroSAT linear-probing evaluation; if accuracy still matches the reported gains, the improvement is a dataset artifact rather than language grounding. A complementary check is to have human annotators rate a random sample of Spectral-Inst captions against the full multispectral images.

Watch

Extended reading notes

Core claim

The paper reports that conditioning a frozen multispectral encoder on language is a representation-enhancement step. After training only the linear projector to match SpectralGPT output tokens to LLaMA3 word embeddings for either BigEarthNet-v2 class labels or ShareCaptioner-generated scene captions, the resulting language-grounded features show tighter t-SNE clusters and higher linear-probe accuracy on EuroSAT across all tested train splits than the raw encoder features. Scene-description alignment yields the best captions after instruction tuning, while class-label alignment is a cheaper signal that still boosts classification. The paper also constructs and releases Spectral-Inst, an instruction dataset of 549,488 multispectral images from BigEarthNet-v2, as the substrate for this alignment.

Load-bearing premise

The English captions from ShareCaptioner, conditioned on RGB-rendered images plus BigEarthNet metadata, are accurate enough to teach the projector semantically correct multispectral concepts; the paper itself concedes the captions are pseudo-data with uncertain accuracy.

Editorial extensions

If this is right

Linear probing on language-grounded features beats raw SpectralGPT features on EuroSAT at every train-test split from 10% to 90%, so text alignment can substitute for visual fine-tuning in classification.
Using scene descriptions in the projector gives the best description quality, while class labels give a cheaper alignment that still helps classification; the choice of alignment text can be matched to the downstream task.
The same frozen encoder can support joint classification and description through task tokens such as [classification] and [caption], so a single model handles both queries.
The higher revised LLaVA-bench detail score (61.6 vs 27.5) indicates that multispectral descriptions are more informative than an RGB-only chat baseline for scenes where RGB is inadequate.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same projector-only alignment could be applied to other frozen remote-sensing encoders, including temporal or SAR models, to obtain vision-language assistants without retraining their backbones.
A natural next test is retrieval: if language grounding genuinely adds semantics, aligned features should also improve image-to-text and text-to-image ranking on multispectral benchmarks.
Because the captions are generated from RGB renderings and metadata rather than from the full spectral content, the size of the reported gain may depend on caption quality; a small human-annotated caption sample would show how much signal survives caption noise.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this paper. First, it does something genuinely useful: it takes the standard LLaVA recipe—frozen SpectralGPT encoder, trainable linear projector, LoRA-tuned LLaMA-3—and applies it to multispectral Sentinel-2 data, contributing a new instruction-tuning dataset (Spectral-Inst) built from BigEarthNet-v2 with ShareCaptioner-generated descriptions. Second, the central quantitative claim that language-grounded alignment improves classification is not actually established by the experiments as reported, and the missing control is simple to run.

The classification experiment in §V-B compares raw SpectralGPT features to aligned features (W·Z_v) via linear probing on EuroSAT. The projector W was trained on BigEarthNet-v2 using class-label strings or ShareCaptioner captions. EuroSAT and BigEarthNet share land-cover semantics (SeaLake, Residential, Pasture, Forest, etc.). So W is a supervised linear map trained on labels closely related to the EuroSAT target classes. A linear probe on top of W·Z_v can improve for a reason that has nothing to do with language: any supervised linear transformation trained on overlapping label structure could reshape features to make EuroSAT classes more linearly separable. The paper reports no control projector trained on BigEarthNet labels without language. That is a load-bearing gap, and it applies even to the class-label alignment, where caption quality is not in question. The stress-test note is right about this.

What the paper does well: the dataset construction is a real contribution, the linear probing protocol is clearly described with 5-fold cross-validation across training splits, and the authors are honest that the captions are pseudo-data with uncertain accuracy. The t-SNE visualization is suggestive but not evidence. The captioning evaluation in Table I rests on an unvalidated self-reported metric on a modified benchmark with no human evaluation or agreement statistics, and the baseline GeoChat is evaluated on RGB inputs, so the comparison is not apples-to-apples. The claimed code and data release is not actually linked in the text.

My bottom line: this is a plausible and reproducible-in-principle recipe, and the Spectral-Inst dataset deserves attention. But the headline attribution of the classification gain to language grounding is unsecured. A serious referee should ask for the non-language control, a real caption evaluation, and the release of artifacts. If the control shows that any supervised BigEarthNet-trained projection gives the same boost, the paper's story shrinks to a dataset-and-recipe contribution—still useful, but not the claim being made. I would send it to review, not desk-reject it, because the core idea is sound and the weaknesses are addressable with additional experiments.

Referee Report

3 major / 5 minor

Summary. The paper proposes Spectral-LLaVA, a vision-language framework in which frozen SpectralGPT multispectral features are linearly projected into the LLaMA-3 embedding space using text-aligned supervision (BigEarthNet-v2 class labels or ShareCaptioner-generated scene descriptions), followed by LoRA fine-tuning of the LLM for scene classification and description generation. The authors introduce Spectral-Inst, an instruction-tuning dataset built from BigEarthNet-v2 with pseudo-captions, and report two sets of experiments: linear probing on EuroSAT comparing raw SpectralGPT features with language-aligned features, and LLaVA-bench-style description scores comparing Spectral-LLaVA with GeoChat. They conclude that language-grounded alignment improves both classification and multispectral scene description.

Significance. The result, if properly supported, would be a convenient recipe for adapting frozen spectral foundation models to instruction-following multimodal LLMs with only a linear projector and LoRA, and the proposed Spectral-Inst dataset could benefit the community. The classification experiments are clearly specified in terms of linear probing, optimizer, and 5-fold cross-validation across train split ratios, and the use of a frozen backbone is a practical advantage. However, the two headline empirical claims currently rest on a missing supervised-transfer control and an unvalidated caption-quality metric, so the significance is conditional on the issues below being resolved.

major comments (3)

[Section V-B, Fig. 4] The claim that language grounding causes the EuroSAT accuracy increase is not established because the projection W is trained on BigEarthNet-v2 labels/descriptions that share land-cover semantics with EuroSAT. Under the class-label variant, W is a linear map fit to BigEarthNet-v2 categories such as SeaLake, Residential, Pasture, and Forest, and EuroSAT categories are nearly the same semantic set. A linear classifier on W Z_v can therefore improve over Z_v simply because W was supervised on overlapping label structure, without any contribution from language. No control is reported that trains W on BigEarthNet-v2 labels with non-linguistic targets, or that applies a BigEarthNet-trained linear head directly to Z_v, so Fig. 4 cannot be read as evidence for language-grounded representation. Please add such controls or explicitly reframe the claim as a supervised-transfer result.
[Section V-A2 Results and Table I] The scene-description evaluation is not described with enough rigor to support the headline improvement. The text states that since no ground-truth explanations exist, descriptions were evaluated by "examining how well they aligned with the class labels" and by criteria such as helpfulness, relevance, accuracy, and detail, yet it does not specify who performed this examination, how the LLaVA-bench detail scores were computed, what the revised benchmark contains, or how many examples were scored. No human evaluation or inter-annotator agreement is reported, and Section IV itself concedes that the ShareCaptioner outputs used to train the projector are "pseudo-data due to its uncertain accuracy." As written, Table I (61.6 vs 27.5) is impossible to interpret as a measure of caption quality; it could reflect label-consistency bias, benchmark selection effects, or an artifact of noisy captions. Please provide the full protocol, release the evaluation prompts and sample, and add human or validated automatic evaluation with confidence intervals.
[Abstract and Section V] The paper claims to demonstrate "substantial improvements through the incorporation of multispectral information," but no controlled RGB baseline of the proposed framework is included. GeoChat is an RGB model but differs from Spectral-LLaVA in backbone, projector, LLM, and training data, so its lower score does not isolate the multispectral contribution. An RGB-only variant of Spectral-LLaVA under the same training and evaluation setup is needed to support the claim that multispectral input is responsible for the improvement.

minor comments (5)

[Section IV] The sentence "With its 549,488 multispectral satellite images captured by Sentinel-2." is a sentence fragment and should be completed.
[Abstract and Section IV] The manuscript states that code and dataset are "available here" but no URL or link is provided; please include a working link.
[References] References [20], [21], [23], [25], [26], and [29] cite arXiv IDs such as "2411.12345", "2401.12345", and "2407.12345" that appear to be placeholders; please verify and replace them with correct identifiers.
[Section V-B, Fig. 4] Please report variance across the five folds, for example with error bars or standard deviations, and specify whether the same folds are used across all train-test split ratios.
[Section I and II] The phrase "first-of-its-kind spectral-domain vision-language framework" should be softened or supported by a more thorough comparison with existing multi-sensor instruction-following models such as EarthGPT/MMRS-1M, which also handle non-RGB modalities.

Circularity Check

2 steps flagged · score 6.0 of 10

Classification gain is confounded with supervised label transfer, and description quality is measured against the same labels used to generate the training captions.

fitted input called prediction [Section V-A1 (Image-Text Alignment) and Section V-B (Scene Classification), Fig. 4]
"Language-Grounded Features Derived from Class Labels: The alignment layer generates language-grounded features by training with multi-label class annotations from the BigEarthNet-v2 dataset. / simple linear probing (a linear neural network layer) is applied on features of vision encoder or language-grounded features to classify given features. / Fig. 4 demonstrates that EuroSAT test data classification performance of language-grounded features ... outperforms vision only SpectralGPT features with a significant margin at every train-test split ratios."

The EuroSAT linear-probe improvement is attributed to 'language-grounded' alignment, but the projector W was fit to BigEarthNet-v2 multi-label land-cover annotations. EuroSAT is also a land-cover classification benchmark with overlapping classes (SeaLake, Residential, Pasture, Forest, etc.), so the gain is expected from any supervised linear projection trained on the same label structure. No non-language control projector (e.g., a linear map trained on BigEarthNet labels without LLM embeddings) is reported, so the result cannot isolate language grounding; the 'prediction' reduces to supervised transfer from the fitted label-conditioned W.
self definitional [Section IV (Generating Spectral Domain Multimodal-Instruction Dataset) and Section V-A2 (Multimodal Finetuning), Table I]
"To improve caption accuracy and semantic richness, we integrate metadata—including image labels and spatial attributes—into the captioning process. / Since ground truth explanation data is unavailable for this domain, we evaluated the generated descriptions by examining how well they aligned with the class labels in the dataset."

The training captions are generated by ShareCaptioner conditioned on BigEarthNet image labels; the description evaluation measures alignment with those same class labels. 'Accurate and detailed description' is therefore operationalized as reproducing the label information that was already an input to caption generation. This makes the description-quality result in Table I partially self-validating rather than an independent test of multispectral semantic understanding.

full rationale

This is not a derivation paper, and there are no equation-level identity derivations or self-citation chains: the visual backbone, projector architecture, and datasets are external and independently cited. The circularity concerns are concentrated in two places. First, the central classification claim (Fig. 4) compares raw SpectralGPT features with features produced by a projector trained on BigEarthNet-v2 multi-label class annotations; because EuroSAT is a land-cover benchmark with semantically overlapping classes, the observed linear-probe gain is consistent with ordinary supervised transfer and is not isolated as an effect of language grounding. Second, the description-quality evaluation uses the same class labels as the target metric that were also fed into the caption-generation process, so the reported description gains are partly constructed by the evaluation protocol. The paper is transparent about the pseudo-data nature of the captions, but that transparency does not remove the confound. Overall, the framework and dataset are real contributions, but the two highlighted 'improvements' are partially circular in their attribution or measurement.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central claim rests on the quality of the generated captions, the pretrained SpectralGPT features, and a self-defined evaluation metric. No new physical or mathematical entities are invented, but Spectral-Inst is a new data artifact whose captions carry noise that is acknowledged by the authors.

free parameters (3)

projection matrix W = learned during alignment
The linear layer mapping SpectralGPT features into LLaMA-3 embedding space is fitted on captions or class labels; it is the core trainable component whose quality determines the reported gains.
LoRA rank r = 128 = 128
Chosen by hand for the LLM fine-tune; the paper gives no ablation or sensitivity analysis.
linear probe hyperparameters (LR 0.0001, 100 epochs, batch 100) = reported in Section V-B
Chosen for the EuroSAT linear probing evaluation; no sensitivity analysis is reported.

assumptions (3)

domain assumption ShareCaptioner captions conditioned on RGB renderings plus BigEarthNet metadata are sufficiently accurate to supervise alignment.
The captioning dataset Spectral-Inst is generated, not human-verified; the authors describe it as pseudo-data in Section IV.
domain assumption SpectralGPT encoder features are a suitable frozen backbone for multispectral representation.
The paper relies on the pretrained SpectralGPT encoder without updating it; no alternative backbone comparison is provided.
domain assumption The revised LLaVA-bench scores are a meaningful proxy for description quality.
The metric is self-assessed with no human evaluation or inter-annotator agreement; the revision procedure is not specified.

invented entities (1)

Spectral-Inst dataset
purpose: Instruction-tuning dataset of multispectral images with generated captions and class labels.
A new dataset artifact whose captions are produced by ShareCaptioner and not validated by humans; its usefulness is only demonstrated indirectly through EuroSAT classification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Vision-Language Framework for Multispectral Scene Representation Using Language-Grounded Features." pith.science (2026). https://pith.science/paper/WZQVQ3Y7

@misc{pith2026250110144,
  author       = {Pith},
  title        = {Pith review of: A Vision-Language Framework for Multispectral Scene Representation Using Language-Grounded Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZQVQ3Y7}},
  note         = {Machine review of arXiv:2501.10144}
}

read the original abstract

Scene understanding in remote sensing often faces challenges in generating accurate representations for complex environments such as various land use areas or coastal regions, which may also include snow, clouds, or haze. To address this, we present a vision-language framework named Spectral LLaVA, which integrates multispectral data with vision-language alignment techniques to enhance scene representation and description. Using the BigEarthNet v2 dataset from Sentinel-2, we establish a baseline with RGB-based scene descriptions and further demonstrate substantial improvements through the incorporation of multispectral information. Our framework optimizes a lightweight linear projection layer for alignment while keeping the vision backbone of SpectralGPT frozen. Our experiments encompass scene classification using linear probing and language modeling for jointly performing scene classification and description generation. Our results highlight Spectral LLaVA's ability to produce detailed and accurate descriptions, particularly for scenarios where RGB data alone proves inadequate, while also enhancing classification performance by refining SpectralGPT features into semantically meaningful representations.

Figures

Figures reproduced from arXiv: 2501.10144 by the authors.

**Figure 2.** Qualitative chat samples with given multispectral images (images in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Test Accuracy Comparison for SpectralGPT and Spectral-LLaVA [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 19 canonical work pages

[1]

Canopy averaged chlorophyll content pre- diction using convolutional autoencoder on hyperspectral data,

S. e. a. Paul, “Canopy averaged chlorophyll content pre- diction using convolutional autoencoder on hyperspectral data,” IEEE JSTARS, vol. 13, pp. 1426–1437, 2020

work page 2020
[2]

Enhancing deforestation monitoring in the brazilian amazon: A semi-automatic approach leveraging uncer- tainty estimation,

J. A. C. Martinez, G. A. O. P. da Costa, C. G. Messias, L. de Souza Soler, C. A. de Almeida, and R. Q. Feitosa, “Enhancing deforestation monitoring in the brazilian amazon: A semi-automatic approach leveraging uncer- tainty estimation,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 210, pp. 110–127, 2024

work page 2024
[3]

Sen1floods11: a georeferenced dataset to train and test deep learning flood algorithms for sentinel-1,

D. Bonafilia, B. Tellman, T. Anderson, and E. Issenberg, “Sen1floods11: a georeferenced dataset to train and test deep learning flood algorithms for sentinel-1,” in IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2020

work page 2020
[4]

Improved base- lines with visual instruction tuning,

H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved base- lines with visual instruction tuning,” arXiv preprint arXiv:2310.03744, 2023

arXiv 2023
[5]

Llava-next: Improved reasoning, ocr, and world knowledge,

H. Liu, C. Li, Y . Li, B. Li, Y . Zhang, S. Shen, and Y . J. Lee, “Llava-next: Improved reasoning, ocr, and world knowledge,” January 2024. [Online]. Available: https://llava-vl.github.io/blog/2024-01-30-llava-next/

2024
[6]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” 2023

2023
[7]

Blip- 3: A family of open large multimodal models,

L. Xue, M. Shu, A. Awadalla, J. Wang et al. , “Blip- 3: A family of open large multimodal models,” arXiv: 2408.08872, 2024

arXiv 2024
[8]

Geochat: Grounded large vision- language model for remote sensing,

K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. Khan, “Geochat: Grounded large vision- language model for remote sensing,” in IEEE Conf. on Computer Vision and Pattern Recognition , 2024

work page 2024

Show all 30 references

[9]

Geollava: Efficient vision-language models for temporal change detection in remote sensing,

H. e. a. Elgendy, “Geollava: Efficient vision-language models for temporal change detection in remote sensing,” arXiv, 2024

2024
[10]

Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding,

J. Luo, Z. Pang, Y . Zhang, T. Wang, L. Wang, B. Dang, J. Lao, J. Wang, J. Chen, Y . Tan, and Y . Li, “Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding,” arXiv preprint arXiv:2406.10100, 2024

2024 arXiv
[11]

reben: Refined bigearth- net dataset for remote sensing image analysis,

K. N. Clasen, L. Hackel, T. Burgert, G. Sumbul, B. Demir, and V . Markl, “reben: Refined bigearth- net dataset for remote sensing image analysis,” arXiv preprint arXiv:2407.03653, 2024

2024 arXiv
[12]

Sharegpt4v: Improving large multi-modal models with better captions,

L. Chen, J. Li, X. Dong, P. Zhang, C. He, J. Wang, F. Zhao, and D. Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” arXiv preprint arXiv:2311.12134, 2023

2023 arXiv
[13]

Spectral reconstruction from satellite multispectral imagery using convolution and transformer joint network,

D. Du, Y . Gu, T. Liu, and X. Li, “Spectral reconstruction from satellite multispectral imagery using convolution and transformer joint network,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, 2023

2023
[14]

Ringmo: A remote sensing foundation model with masked image modeling,

X. Sun, P. Wang, W. Lu, Z. Zhu et al. , “Ringmo: A remote sensing foundation model with masked image modeling,” IEEE Transactions on Geoscience and Re- mote Sensing, 2022

2022
[15]

Masked autoencoders are scalable vision learners,

K. He, X. Chen, S. Xie, Y . Li, P. Doll´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” in IEEE Conf. on Computer Vision and Pattern Recognition, 2022

2022
[16]

Prithvi- eo-2.0: A versatile multi-temporal foundation model for earth observation applications,

D. Szwarcman, S. Roy, P. Fraccaro et al. , “Prithvi- eo-2.0: A versatile multi-temporal foundation model for earth observation applications,” arXiv preprint arXiv:2412.01922, 2024

2024 arXiv
[17]

Spectralgpt: Spectral remote sensing foundation model,

D. Hong, B. Zhang, X. Li et al. , “Spectralgpt: Spectral remote sensing foundation model,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5227–5244, 2024

2024
[18]

Rsgpt: A remote sensing vision-language model and benchmark,

Y . e. a. Hu, “Rsgpt: A remote sensing vision-language model and benchmark,” arXiv, 2023

2023
[19]

Earthmarker: A visual prompting multi-modal large language model for remote sensing,

W. Zhang, M. Cai, T. Zhang, Y . Zhuang, J. Li, and X. Mao, “Earthmarker: A visual prompting multi-modal large language model for remote sensing,”IEEE Transac- tions on Geoscience and Remote Sensing , pp. 1–1, 2024

2024
[20]

Rs-moe: Mixture of experts for remote sensing image captioning and visual question answering,

H. Lin, D. Hong, S. Ge, C. Luo, K. Jiang, H. Jin, and C. Wen, “Rs-moe: Mixture of experts for remote sensing image captioning and visual question answering,” arXiv preprint arXiv:2411.12345, 2024

2024 arXiv
[21]

Lhrs-bot-nova: Improved multimodal large language model for remote sensing vision-language in- terpretation,

Z. Li, D. Muhtar, F. Gu, X. Zhang, P. Xiao, G. He, and X. Zhu, “Lhrs-bot-nova: Improved multimodal large language model for remote sensing vision-language in- terpretation,” arXiv preprint arXiv:2411.09301 , 2024

2024 arXiv
[22]

Geoground: A unified large vision-language model. for remote sensing visual grounding,

Y . Zhou, M. Lan, X. Li, Y . Ke, X. Jiang, L. Feng, and W. Zhang, “Geoground: A unified large vision-language model. for remote sensing visual grounding,” 2024. [Online]. Available: https://arxiv.org/abs/2411.11904

2024 arXiv
[23]

Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tun- ing with large language model,

Y . Zhan, Z. Xiong, and Y . Yuan, “Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tun- ing with large language model,” 2024

2024
[24]

Ringmogpt: A unified remote sensing foundation model for vision, language, and grounded tasks,

P. Wang, H. Hu, B. Tong, Z. Zhang et al., “Ringmogpt: A unified remote sensing foundation model for vision, language, and grounded tasks,” IEEE Transactions on Geoscience and Remote Sensing , pp. 1–1, 2024

2024
[25]

Teochat: A vision-language assistant for earth observation data,

J. A. e. a. Irvin, “Teochat: A vision-language assistant for earth observation data,” arXiv, 2024

2024
[26]

The llama 3 herd of models,

A. Grattafiori, A. Dubey, A. Jauhri et al., “The llama 3 herd of models,” arXiv preprint arXiv:2407.12345, 2024

2024 arXiv
[27]

Functional map of the world,

G. Christie, N. Fendley, J. Wilson, and R. Mukherjee, “Functional map of the world,” in CVPR, 2018

2018
[28]

BigEarthNet: A large-scale benchmark archive for re- mote sensing image understanding,

G. Sumbul, M. Charfuelan, B. Demir, and V . Markl, “BigEarthNet: A large-scale benchmark archive for re- mote sensing image understanding,” in IGARSS. IEEE, 2019, pp. 5901–5904

2019
[29]

Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain,

W. Zhang, M. Cai, T. Zhang, Y . Zhuang, and X. Mao, “Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain,” arXiv preprint arXiv:2401.12345 , 2024

2024 arXiv
[30]

Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,

P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,” IEEE Journal of Se- lected Topics in Applied Earth Observations and Remote Sensing, 2019

2019

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Canopy averaged chlorophyll content pre- diction using convolutional autoencoder on hyperspectral data,

S. e. a. Paul, “Canopy averaged chlorophyll content pre- diction using convolutional autoencoder on hyperspectral data,” IEEE JSTARS, vol. 13, pp. 1426–1437, 2020

work page 2020

[2] [2]

Enhancing deforestation monitoring in the brazilian amazon: A semi-automatic approach leveraging uncer- tainty estimation,

J. A. C. Martinez, G. A. O. P. da Costa, C. G. Messias, L. de Souza Soler, C. A. de Almeida, and R. Q. Feitosa, “Enhancing deforestation monitoring in the brazilian amazon: A semi-automatic approach leveraging uncer- tainty estimation,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 210, pp. 110–127, 2024

work page 2024

[3] [3]

Sen1floods11: a georeferenced dataset to train and test deep learning flood algorithms for sentinel-1,

D. Bonafilia, B. Tellman, T. Anderson, and E. Issenberg, “Sen1floods11: a georeferenced dataset to train and test deep learning flood algorithms for sentinel-1,” in IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2020

work page 2020

[4] [4]

Improved base- lines with visual instruction tuning,

H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved base- lines with visual instruction tuning,” arXiv preprint arXiv:2310.03744, 2023

arXiv 2023

[5] [5]

Llava-next: Improved reasoning, ocr, and world knowledge,

H. Liu, C. Li, Y . Li, B. Li, Y . Zhang, S. Shen, and Y . J. Lee, “Llava-next: Improved reasoning, ocr, and world knowledge,” January 2024. [Online]. Available: https://llava-vl.github.io/blog/2024-01-30-llava-next/

2024

[6] [6]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” 2023

2023

[7] [7]

Blip- 3: A family of open large multimodal models,

L. Xue, M. Shu, A. Awadalla, J. Wang et al. , “Blip- 3: A family of open large multimodal models,” arXiv: 2408.08872, 2024

arXiv 2024

[8] [8]

Geochat: Grounded large vision- language model for remote sensing,

K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. Khan, “Geochat: Grounded large vision- language model for remote sensing,” in IEEE Conf. on Computer Vision and Pattern Recognition , 2024

work page 2024