acceptodds
Withdrawn from ICLR 2027

SatLAERS: A Resource-Efficient Remote Sensing VLM for Onboard Disaster Understanding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Efficient remote sensing vision-language models (RS-VLMs) can enable onboard first-look disaster understanding from the latest available high-resolution optical observation, providing multi-granularity information for rapid response. Recent RS-VLMs have made substantial progress in unified multi-task understanding and model efficiency, yet disaster-oriented models remain comparatively limited and are often specialized to individual tasks. We introduce SatLAERS, a granularity-aware adaptive RS-VLM for resource-efficient adaptation and onboard inference. Built on a frozen vision encoder and compact language backbone, SatLAERS allocates multimodal computation according to task granularity and sample complexity, combining lightweight multi-scale visual adaptation and query-adaptive visual routing with sparse low-rank experts. A progressive three-stage pipeline proceeds from efficient vision-language alignment and multi-granularity instruction tuning to budget-aware reinforcement learning, jointly balancing task performance and computational cost. We further introduce SatDISA, a single-observation disaster benchmark with 22,862 high-resolution images and 316,698 vision-language instructions across six disaster types and four semantic granularities, spanning image-level understanding, region-level interpretation, pixel-level grounding, and disaster-oriented situational understanding. SatLAERS completes domain adaptation on 8 NVIDIA RTX 3090 GPUs in 49.3 GPU-hours, outperforms the strongest comparable-scale adapted baseline by 1.55 points, and achieves 16.99 s end-to-end latency with 4.461 GB peak RAM on Jetson Orin NX Super, supporting resource-constrained onboard deployment.

Primary area:
foundation or frontier models, including LLMs
Keywords:
vision-language models, remote sensing, efficient multimodal learning, onboard inference, disaster understanding

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1EWHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SatLAERS for disaster understanding from a single remote sensing image under limited onboard computing resources. It keeps the vision and language backbones frozen and learns lightweight modules that select visual tokens and activate experts according to the query, a reinforcement-learning stage balances task scores against token usage and response length. It also introduces the SatDISA disaster benchmark.

Critical strengths

The controlled ablation study is useful. This supports a potentially useful accuracy–computation trade-off within the authors' setup. The overall performance of the proposed model is good, when compared with other models (in Table 1).

The paper includes hardware measurements rather than parameter counts alone. Table 5 (p. 9) reports 16.99 s latency and 4.461 GB memory on a Jetson device. This is relevant evidence for an edge prototype.

Critical weaknesses

The benchmark does not establish independent generalization. SatDISA is intended for “systematic evaluation (p. 4)” , but the paper gives no train/validation/test partition or scene/event grouping. With many instructions per image and overlapping disaster sources, instruction-level separation would not establish unseen-scene or unseen-event performance. Leakage is not demonstrated. The problem is that the reported score cannot be interpreted as independent disaster understanding without a defined evaluation population.

The experiments do not isolate an efficiency benefit from adaptive computation. The paper attributes the trade-off to adaptive routing, but compares systems with different backbones and adaptation procedures. Its existing fixed-budget/no-RL controls have no hardware measurements. Token retention and response length are proxies, not equivalent to measured execution cost. Table 5 shows SatLAERS is slower and uses slightly more memory than Qwen3-VL.

The value of handling all disaster tasks through one VLM need to be evaluated. The paper emphasizes a "unified autoregressive framework" (p. 4), but Table 1 compares only VLM-based systems. There is no task-specific detection or segmentation reference to establish what spatial accuracy is gained or sacrificed for the unified interface. For example, the paper reports a segmentation mIoU of 18.76, but how do classical segmentation models (such as UperNet) perform? Furthermore, an mIoU of 18.76 does not appear to be of practical value in real-world applications.

Questions for authors

See above

Reviewer_EyndOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SatLAERS is a small vision-language model, about two billion parameters, for understanding disaster scenes from a single satellite image on board. It freezes a pretrained image encoder and language model and trains light add-ons that view the image at three scales, keep only as many image tokens as a question seems to need, and add small adapters inside the language model. Training ends with reinforcement learning that rewards correct answers and penalizes token use and long outputs. The paper also builds SatDISA, about 23,000 images and 317,000 instructions over six disaster types, with tasks from naming the disaster to locating and outlining affected areas and judging risk, assembled from xBD and several other public datasets. SatLAERS scores 57.31 overall against 55.76 for the best fine-tuned baseline, trains in about 49 GPU hours, and answers a query in about 17 seconds with 4.5 GB of memory on an NVIDIA Jetson Orin NX.

Critical strengths

The paper measures cost on real edge hardware (first-token latency, end-to-end latency, peak memory) and reports token use and output length next to the scores in every ablation, which few RS-VLM papers do. Table 3 shows a sensible learned allocation, about 0.29 of the visual tokens for classification and 0.68 for pixel tasks.

Critical weaknesses
  1. Training is cheaper (Table 4), but Table 5 does not support the onboard inference claim. SatLAERS has the slowest first token of the three models (555 ms against 315 ms for Qwen3-VL-2B), is about 50% slower end to end than Qwen3-VL-2B (16.99 s against 11.27 s), and uses slightly more memory. Section 5.4 calls this a trade-off, but making inference cheaper is what the router, the budget policy and the RL cost terms are for. A likely cause is the adapter, whose three scales give 1,344 visual tokens, more than the finest 32×32 grid alone, so about 620 still reach the language model at 46% retention. Two runs are missing, SatLAERS without routing, and a compact detector and segmenter on the same Jetson. The second is the real onboard competitor, and against 31.26 on detection and 18.76 mIoU on segmentation I would expect it to win at a fraction of the latency.
  2. SatDISA is half the contribution and every result is measured on it, yet its split and label generation are not described. There are no train and test sizes, nothing on whether test images come from events seen in training (xBD puts tiles of the same event on both sides), and no account of how the instructions and the risk, severity and action targets were made, by template or by the eight experts. Sources may also confound the classes. Landslides come from the three landslide datasets the paper names and wildfire partly from a separate wildfire set, with different sensors and resolutions, so the 83.59% classification accuracy may partly be source recognition. No data or code release is mentioned, and the appendix is the empty template.
Questions for authors
  1. How is the reasoning length in Table 3 counted? One 32-point polygon already holds 64 coordinates, yet pixel tasks average 27.
  2. Without the RL stage SatLAERS is only 0.79 above tuned TinyRS-R1, which had no RL stage (Table 4). What do the baselines reach with comparable RL?
  3. The prompts state the task type outright. How does the learned budget compare with a fixed budget per task type?
  4. Is detection [email protected] reported as precision, recall, AP or matched accuracy?
  5. Please give the cost weights, minimum budget, expert rank and count, GRPO group size and input resolution, what the 16.99 s covers, and whether the Qwen3-VL-2B row in Table 5 is the tuned model.
  6. Will SatDISA be released, and under what license given the xBD and Maxar Open Data terms?
  7. The 9.07-point margin over off-the-shelf models partly reflects unfamiliar box and polygon formats and could go.
  8. Region description BERTScore is above 82 for every model, which flattens the region-level average.
Reviewer_n6iVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SatLAERS, a resource-efficient remote sensing vision-language model for onboard disaster understanding from single satellite observations. The framework integrates multi-scale visual adaptation, query-adaptive visual token routing, and sparse low-rank experts, trained through a three-stage pipeline involving vision-language alignment, instruction tuning, and budget-aware reinforcement learning. The authors also introduce SatDISA, a multi-granularity disaster understanding benchmark covering six disaster types.

Critical strengths
  1. The paper addresses an important problem of resource-efficient disaster understanding from satellite imagery, where computational and memory constraints are particularly relevant for onboard applications.
  2. SatLAERS integrates image-level recognition, region-level interpretation, pixel-level grounding, and situational understanding within a single framework.
  3. The combination of multi-scale visual adaptation, query-adaptive token selection, and sparse low-rank experts is well motivated.
Critical weaknesses
  • The proposed components largely build upon existing techniques, and the paper does not clearly distinguish its methodological contributions from prior adaptive computation frameworks.
  • Although RSUniVLM and SmartVL are discussed in the related work, neither is included in the experimental comparisons.
  • All experiments are conducted on SatDISA, without evaluating generalization to unseen disaster datasets or geographic regions.
  • The reported 1.55-point improvement over the strongest adapted baseline is relatively small, but no standard deviations or multiple-run results are provided.
  • The paper does not adequately explain how multiple bounding boxes are matched or how invalid predictions are handled when computing [email protected].
  • Visual token selection occurs after feature extraction, and the reported token reduction is not accompanied by FLOPs or energy measurements.
  • Although Table 2 shows reduced token consumption after GRPO, the paper does not report corresponding inference latency improvements.
  • Figure 3 includes disaster spread-direction inference from a single satellite image, although such predictions generally require temporal or environmental information.
Questions for authors

Please see the Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.