PaperlayerView exact revision on arXiv

Source-linked AI summary

GuidedSAM-Plume: An Expert-Guided Segmentation Framework for Volcanic Plumes

Taddeo D'Adamo, Simon Thivet, Catherine Pothier, Riccardo Simionato, Allan Fries, Jonathan Lemus, Costanza Bonadonna, Laure Tougne Rodet, Bertrand Kerautret

arXiv:2610.29825v1cs.CV

TL;DR

Volcanic plume analysis lacks dedicated segmentation solutions, and partly handcrafted annotations remain common. GuidedSAM-Plume applies zero-shot SAM2 with expert-guided interactive refinement and is reported to improve segmentation over fully automatic methods while preserving the reliability of derived measurements. The work also releases an expert-guided re-annotated dataset and interface.

  • Problem

    Dedicated segmentation solutions for volcanic imagery are scarce, while partly handcrafted masks require effort and can vary across operators, complicating quantitative plume characterization.

    Source: §1 · pp. 1–2

  • Method

    GuidedSAM-Plume applies SAM2’s zero-shot spatial prompting in an interactive, expert-guided framework for plume segmentation.

    Source: §1 · p. 2

  • Results

    The full expert-guided framework shows considerable improvement over fully automatic methods while preserving the reliability of derived volcanological measurements.

    Source: Quantitative analysis · p. 9, §6 · p. 12

  • Takeaways & Limitations

    The work releases an expert-supervised re-annotated dataset with more standardized masks and makes both annotation versions publicly available.

    Source: §1 · p. 2

    The modest dataset and relatively short temporal sequences narrow the scope of quantitative evaluation, and the framework still requires expert intervention.

    Source: §6 · p. 13

Abstract

from arXiv · show

Quantitative analysis of environmental phenomena increasingly relies on image-based measurements. In volcanology, the geometric characterization of volcanic plumes and clouds from RGB images remains hindered by the scarcity of dedicated segmentation solutions and partially handcrafted annotations remain widely prevalent. We present GuidedSAM-Plume, an open-source expert-guided framework tailored for the interactive segmentation of volcanic plumes and clouds. It leverages the zero-shot, spatial prompt-based capabilities of SAM2. The framework was thoroughly validated on a challenging dataset through dedicated segmentation metrics and field-relevant volcanological descriptors. From minimal user inputs, it is able to significantly increase performance compared to fully automatic dedicated solutions while preserving the reliability of derived volcanological parameters. The absence of retraining requirements lowers adoption barriers within the volcanological community, and the framework's demonstrated suitability for complex, diffuse plume-shaped phenomena extends its applicability to broader environmental science domains. The re-annotated dataset, produced under the guidance of expert volcanologists within the proposed framework, ultimately yields a more standardized release. The framework and accompanying dataset are made available at https://github.com/tad-damo/GuidedSAM-Plume.

1 Introduction

Volcanic plume segmentation remains limited by scarce dedicated solutions and reliance on partly handcrafted masks, despite its importance for quantitative analysis. GuidedSAM-Plume combines zero-shot SAM2 with expert guidance, and the work also releases a re-annotated dataset validated for downstream volcanological analysis.

Motivation

  • Plume delineation supports quantitative analysis of morphology, spatial extent, and temporal dynamics, but dedicated solutions remain scarce and handcrafted masks can introduce annotation effort and cross-dataset inconsistencies.The paper identifies operator-driven inconsistencies as a source of subjective bias when comparing studies.Source: §1 · p. 2

Contributions

  • GuidedSAM-Plume uses SAM2’s zero-shot capabilities in an expert-in-the-loop workflow for reproducible image and video segmentation while reducing annotation costs without extensive retraining.The framework was developed through quantitative analysis and close collaboration with volcanologists to address operational needs.Source: §1 · p. 2
  • The released dataset contains 47 observational sequences and 759 annotated frames, re-annotated under expert supervision to produce masks described as more standardized than those from independent annotators.Both annotation versions are publicly available; segmentation quality was assessed against the manually annotated ground truth and through application-driven plume descriptors.Source: §1 · pp. 2–3
  • The framework and dataset are publicly available.The paper provides the project repository at https://github.com/tad-damo/GuidedSAM-Plume.Source: §1 · p. 3

2 Related Work

Existing volcanic-plume segmentation approaches include manual, site-specific, and learned systems, while variable imaging conditions and irregular plume structures challenge fully automated methods. Interactive segmentation and foundation models offer a basis for expert-guided workflows, including SAM2’s support for video.

2.1 Volcanology-dedicated solutions

  • Volcanology-specific tools and learned systems remain limited, while variable imaging conditions and irregular plume structures challenge fully automated static-image segmentation.Earlier semi-automatic tools require substantial manual effort, and some deep-learning approaches target individual volcanoes.Source: §2.1 · p. 3

2.2 Interactive Segmentation

  • Interactive segmentation uses iterative refinement to balance annotation effort and output quality, while SAM enables zero-shot prompting and SAM2 adds video segmentation through memory components.SAM accepts prompts such as boxes, points, or coarse masks; SAM2 adds a memory bank and memory attention module.Source: §2.2 · pp. 3–4
  • For still images, GuidedSAM-Plume starts with a bounding box and refines with foreground and background points; for video, it starts on a representative middle keyframe before tracking and refinement across additional keyframes.The figure presents the framework’s workflow as an application of spatial prompting and SAM2 video segmentation.Source: Figure 1 · p. 4, §2.2 · p. 4

3 GuidedSAM-Plume

GuidedSAM-Plume uses spatial prompts and iterative expert refinement to segment still images and sequences. Its dedicated interface supports keyframe-based tracking, image handling, and export.

Interactive workflow

  • Experts initialize segmentation with a bounding box and refine it using positive points inside the plume and negative points in the background, adding sparse prompts until the mask captures the observed geometry.Prompts can be combined; for sequences, further refinement on additional keyframes can improve tracking.Source: §3 · pp. 4–5
  • For time series or video, the recommended workflow begins on a representative frame near the sequence midpoint, then uses temporal tracking and additional keyframes for refinement.For single images, each frame requires user interaction; sequence tracking reduces manual input.Source: §3 · pp. 4–5

Workflow interface

  • The interface workflow includes keyframe prompting for sequences, mask propagation across frames, individual-image handling, and data export.Source: Figure 2 · p. 5, §3 · p. 6
  • The dedicated interface supports segmentation of still images and sequences, integration of multiple prompt types, export, and volcanology-specific metric computation.The interface is built upon the open-source SAM2 Demo codebase.Source: §3 · p. 5

4 Evaluation Method

The evaluation protocol benchmarks SAM2 and existing plume-segmentation solutions, then assesses prompt workflows, segmentation metrics, and volcanological descriptors. It uses a 47-sequence dataset spanning varied plume conditions.

4.1 Protocol

  • The protocol benchmarks SAM2 against Wilkes et al. and VigiaPlume-Net, with expert-guided final masks also evaluated.SAM2 uses its largest 2.1 variant, selecting the highest-confidence of three outputs; the protocol also estimates resolution-related error by downscaling and upscaling ground truths to match SAM2 processing.Source: SAM2 assessment · p. 6
  • Static-image tests compare box and point prompts, random and spatially predetermined points, and up to five prompts per image.Random-prompt metrics are averaged over five draws; video tests assess propagation from manually annotated keyframes and identify a prompting strategy.Source: SAM2 assessment · pp. 6–7

4.2 Metrics

  • Standard segmentation metrics derive from true-positive, true-negative, false-positive, and false-negative counts, including accuracy, precision, recall, Dice, Jaccard, contour F1, and J&F.J&F averages region similarity J and contour measure F.Source: §4.2 · p. 7
  • Volcanological assessment measures aspect ratio, convexity, and solidity, then reports their agreement with ground truth as complements of absolute relative error.The descriptors use Feret diameters, convex-hull perimeter and area, and mask perimeter and area.Source: §4.2 · p. 7

4.3 Dataset

  • Figure 3 presents three keyframes from each of four distinct sequences.Source: Figure 3 · p. 8
  • The dataset contains 47 manually segmented time-series sequences totaling 759 frames, with a mean sequence length of 16 frames.Images average 2996 × 1685 pixels and cover varied viewing angles, illumination, occlusions, weather, and plume-development states; a volcanologist-guided re-annotated version is also released.Source: §4.3 · pp. 7–8

5 Results

Expert-guided refinement improves segmentation over fully automatic methods, while the results characterize its boundary-smoothing limits and practical prompting strategies. The framework also supports consistent delineation of ambiguous scenes and reliable downstream volcanological measurements.

5.1 Comparison with fully automatic methods

  • Targeted expert refinement substantially improves SAM2 segmentation over fully automated methods, supporting the additional annotation effort for quantitative volcanological studies.Even without refinement, SAM2 considerably outperforms existing solutions; VigiaPlume-Net leads in accuracy and precision but underperforms on recall and volcanological metrics.Source: Quantitative analysis · pp. 8–9
  • In challenging scenes, Wilkes et al. produce spurious background detections and VigiaPlume-Net staircase artifacts, while GuidedSAM-Plume standardizes ambiguous delineations.Diffuse boundaries, heterogeneous opacity, small plumes, and out-of-focus imagery can make human annotations vary.Source: Qualitative analysis of challenging… · p. 9, Figure 4 · p. 9

5.2 Maximal expected performance

  • Downscaling and re-upscaling ground truths estimates the architectural floor; edge-sensitive F, J&F, and CVXϵ show the largest losses, while most metric losses are negligible.Contour, convex-hull perimeter, and area errors show a consistently negative bias, reflecting intrinsic SAM2 smoothing.Source: §5.2 · pp. 9–10, Table 2 · p. 10

5.3 Workflow establishment

  • A bounding box consistently beats a single positive point; without one, performance declines even with four to five points, and selective background points are preferred.Adding randomly placed points to a box yields no significant improvement; excessive prompts may be counterproductive.Source: Prompts efficiency · pp. 10–11, Table 3 · p. 11
  • Propagation from ground-truth keyframes is relatively stable; middle-frame initialization has a slight advantage, and adding subsequent keyframes further improves performance.Starting from the first frame performs worse, likely because the plume has limited spatial extent early in the sequence.Source: Video propagation assessment · pp. 11–12, Table 4 · p. 11
  • With the full workflow, all volcanological-parameter metrics exceed 90%, although perimeter and area show predominantly negative deviations consistent with inward boundary smoothing.The deviations remain proportionally small, particularly for area estimates, which are less sensitive to contour errors.Source: Impact on volcanological applications · p. 12

6 Discussion and conclusion

Expert annotation time fell substantially while segmentation agreement and derived volcanological measurements remained reliable. The authors note that broader validation of second-order parameters would help confirm operational robustness.

6 Discussion and conclusion

  • Expert annotation fell from approximately 2 hours to 5–10 minutes per 10-frame sequence, while segmentation maintained strong agreement and reliable derived measurements.Annotation time varied with image resolution and scene complexity.Source: §6 · p. 12
  • Perimeter and area deviations relative to manual masks remained controlled and predominantly systematic in direction.Correcting these biases in downstream analyses is identified as a direction for improving measurement accuracy.Source: §6 · p. 13, Figure 5 · p. 13
  • Further validation of real-scale plume dimensions and dynamic properties would help confirm robustness for operational applications.Source: §6 · p. 13
  • Refinement remains bounded by SAM model input limits and expert intervention, while modest dataset size and short sequences narrow quantitative evaluation.Source: §6 · p. 13
  • Volcanologists deployed the framework in real-world settings to produce the standardized re-annotated dataset.Source: §6 · p. 13
  • Stable performance across diverse plume morphologies and challenging imaging conditions suggests potential beyond volcanology, including dust, clouds, wildfire smoke, and industrial emissions.Source: §6 · pp. 13–14

Related papers

Loading 2610.29825v1…