Research map / Detection and segmentation / Open-set and few-shot detection
Hotspotter: A Generalizable Pipeline for Automated Detection of Subtle Volcanic Thermal Features in Satellite Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Hotspotter, a pipeline that combines unsupervised LRX anomaly scoring with a shallow supervised CNN filter to detect subtle volcanic thermal features in ASTER satellite images and derive thermal statistics.
Problem
Automated detection of subtle volcanic thermal features (VTFs) in high-resolution thermal satellite imagery across a globally diverse set of volcanoes, where labeled data is limited and manual curation is labor-intensive.
Approach
Use an unsupervised Local Reed-Xiaoli (LRX) anomaly detector to produce pixel-wise anomaly score images, then a shallow 3-layer CNN to classify each score image as true VTF or false positive, and extract thermal statistics from the top-scoring pixel.
Key results
- LRX+CNN achieved 90.3% test accuracy and 0.901 F1 on a random split, versus 83.1% accuracy for CNN-only and 63.1% for a fixed threshold.
- On non-overlapping volcano splits, LRX+CNN reached 88.9% accuracy versus 63.7% for CNN-only.
- For unseen volcanoes, thermal statistic MAE was 3.23 K (max temperature), 4.27 K (mean background), 11.48 K (max above background), and 5.16 K (std dev).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LRX anomaly scoring before CNN classification (ablation) | Hotspotter · VTF detection, random split | accuracy: 0.831 -> 0.903 | Helped |
| Add LRX anomaly scoring before CNN classification (ablation) | Hotspotter · VTF detection, non-overlapping split | accuracy: 0.637 -> 0.889 | Helped |
| Use LRX+CNN instead of fixed temperature thresholding | VTF detection · VTF detection, random split | accuracy: 0.631 -> 0.903 | Helped |
| Use a shallow 3-layer CNN rather than deeper configurations (ablation) | Hotspotter CNN classifier · VTF detection | performance | Helped |
Surprising findings
- The CNN-only model performed far worse on non-overlapping volcano splits (63.7%) than on random splits (83.1%), indicating poor generalization without LRX preprocessing.
- A simple 2 K threshold baseline reached only 63% accuracy, below the learned pipeline.
Limitations
- Relies on a single labeled database (Gomez-Patron et al. 2023) for ground truth, and ASTER data is cloud-filtered manually.
- Thermal statistic errors are large for max temperature above background (11.48 K MAE).
- Performance varies across random seeds and train/test splits.
Experimental setup
Datasets: ASTER AST_08 thermal images (2000-2022, 129 volcanoes) Gomez-Patron et al. (2023) VTF database AVTOD
Benchmarks: Hotspotter ASTER VTF test split (7,968 points, 129 volcanoes)
Metrics: accuracy F1 score precision at N mean absolute error (K)
Baselines: CNN-only on thermal images fixed temperature threshold (2 K above background) MOUNTS Corradino et al. U-Net Corradino et al. random forest
Builds on: DORA AVTOD MOUNTS Local Reed-Xiaoli detector
Closest papers by meaning
- FireScope: Wildfire Risk Raster Prediction With a Chain-of-Thought Oracle (CVPR 2026 · similarity 0.80)
- MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting (CVPR 2025 · similarity 0.78)
- Alberta Wells Dataset: Pinpointing Oil and Gas Wells from Satellite Imagery (ICML 2025 · similarity 0.78)
- Towards Automated Petrography (NeurIPS 2025 · similarity 0.78)
- Mars-Bench: A Benchmark for Evaluating Foundation Models for Mars Science Tasks (NeurIPS 2025 · similarity 0.78)
- REOBench: Benchmarking Robustness of Earth Observation Foundation Models (NeurIPS 2025 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.