Source-linked AI summary
Hyperspectral Remote Sensing Benchmark Database for Oil Spill Detection with an Isolation Forest-Guided Unsupervised Detector
Puhong Duan, Xudong Kang, Pedram Ghamisi
TL;DR
Oil spills pose severe environmental and coastal risks, while existing hyperspectral methods depend on costly high-quality labels. This paper proposes an unsupervised isolation-forest pipeline and reports that it outperforms advanced alternatives on HOSD.
Problem
Existing HSI oil-spill methods rely on sufficient high-quality labels that are costly to collect amid spectral variability, while spills threaten marine ecosystems and coastal resources.
Source: §1 · pp. 2, 4Method
The method removes severely noisy bands, uses isolation forest to generate pseudo-labeled samples, trains an SVM, and refines its detection map with ERW.
Source: §1 · p. 4Results
On HOSD, the approach outperformed advanced methods including high-dimensional data clustering, unsupervised object detection, and supervised oil-spill detection.
Source: §6 · p. 20Takeaways & Limitations
HOSD is freely available for oil-spill detection or for training deep models.
Source: §6 · p. 20The automatically generated training samples are not guaranteed to be highly accurate.
Source: §6 · p. 20
Abstract
from arXiv · showhide
Oil spill detection has attracted increasing attention in recent years since marine oil spill accidents severely affect environments, natural resources, and the lives of coastal inhabitants. Hyperspectral remote sensing images provide rich spectral information which is beneficial for the monitoring of oil spills in complex ocean scenarios. However, most of the existing approaches are based on supervised and semi-supervised frameworks to detect oil spills from hyperspectral images (HSIs), which require a huge amount of effort to annotate a certain number of high-quality training sets. In this study, we make the first attempt to develop an unsupervised oil spill detection method based on isolation forest for HSIs. First, considering that the noise level varies among different bands, a noise variance estimation method is exploited to evaluate the noise level of different bands, and the bands corrupted by severe noise are removed. Second, kernel principal component analysis (KPCA) is employed to reduce the high dimensionality of the HSIs. Then, the probability of each pixel belonging to one of the classes of seawater and oil spills is estimated with the isolation forest, and a set of pseudo-labeled training samples is automatically produced using the clustering algorithm on the detected probability. Finally, an initial detection map can be obtained by performing the support vector machine (SVM) on the dimension-reduced data, and then, the initial detection result is further optimized with the extended random walker (ERW) model so as to improve the detection accuracy of oil spills. Experiments on airborne hyperspectral oil spill data (HOSD) created by ourselves demonstrate that the proposed method obtains superior detection performance with respect to other state-of-the-art detection approaches.
1. Introduction
Oil spills threaten marine ecosystems, coastal resources, and human health, while hyperspectral detection faces costly annotation and band-dependent noise. The paper proposes an unsupervised detector and introduces HOSD as a public benchmark dataset.
1. Introduction
- Existing HSI oil-spill methods rely on sufficient high-quality labels that are costly to collect amid spectral variability, while severe band noise also degrades detection.Spectral variability is associated with sun glint, oil thickness, and emulsification stage.Source: §1 · p. 4
- The proposed unsupervised pipeline removes severely noisy bands, uses isolation forest to generate pseudo-labeled samples, then trains an SVM and refines its map with ERW.The method is presented as reducing manual labeling and handling noisy input bands.Source: §1 · p. 4
- HOSD is introduced as a publicly available oil-spill benchmark and, to the authors’ knowledge, the first public dataset of its kind and the largest by total image count.The authors state that the dataset can support advances in oil-spill detection algorithms.Source: §1 · p. 4
- Removing bands contaminated by serious noise is reported to benefit oil-spill detection.The Gaussian statistical method estimates the noise level of each spectral band before severe-noise bands are removed.Source: §1 · p. 5
2. Methodology
The method automatically removes severely noisy bands, reduces spectral dimensionality, and uses isolation-forest probabilities to create training samples for an SVM. Extended random walker refinement then incorporates spatial information into the initial detection map.
2.1. Noisy band removal and dimension reduction
- Gaussian noise estimates and thresholding remove severely noisy bands, while RBF-kernel KPCA reduces the remaining spectral dimension.The Laplacian mask estimates local noise; bands below the noise-variance threshold are preserved.Source: §2.1 · pp. 6–7
2.2. Generation of training samples
- Isolation-forest pixels with shorter average paths yield probabilities that k-means clusters into oil-film and seawater training samples.The method uses spectral differences between oil film and surrounding seawater to motivate isolation-forest probability estimation.Source: §2.2 · pp. 7–9
2.3. Oil spill detection and optimization
- An SVM trained on 1% of the generated samples produces the initial detection map, with five-fold cross-validation used to select its parameters.The 1% sample selection is used to avoid high computing cost.Source: §2.3 · p. 9
- ERW refines the SVM map by combining spatial correlations among neighboring pixels with initial probabilities, reducing noisy pixels and aligning boundaries.The ERW objective has spatial and aspatial terms.Source: §2.3 · p. 9
3. Data sets
The study focuses on the Gulf of Mexico and introduces HOSD, a large-scale airborne hyperspectral oil spill database described as the first public dataset of its kind.
3.1. Study area
- The study area is in the Gulf of Mexico near 25°N, 90°W, where the 2010 Deepwater Horizon disaster released more than 200 million gallons of crude oil.The paper identifies oil-spill detection as relevant to monitoring and cleanup in this area.Source: §3.1 · p. 10
3.2. HOSD database
- HOSD is presented as the first public oil spill detection dataset, with wide distribution, large coverage, and the largest number of data among datasets cited by the authors.Its AVIRIS images span 365–2500 nm, have route-dependent spatial resolution, and use pixelwise expert-guided reference maps; the data were atmospherically corrected with FLAASH.Source: §3.2 · pp. 10–11, Table 1 · p. 11
4. Experiments
Experiments compare the proposed detector with clustering, unsupervised classification, object-detection, and supervised methods on HOSD. The proposed method achieves the highest detection accuracy overall, with parameter analysis selecting D = 25 and T = 800.
4. Experiments
- The comparison includes clustering, unsupervised classification, object-detection, and supervised methods, evaluated using AUC and detection precision (DP).The authors adopt the comparison methods’ recommended parameters where possible.Source: §4 · pp. 11–12
4.1. Parameter analysis
- D = 25 and T = 800 yield the highest AUC in the tested parameter sweep, with a relatively high DP value.D ranged from 5 to 50 in steps of 5, and T from 200 to 1600 in steps of 200.Source: Figure 3 · p. 12, §4.1 · p. 13
4.2. Detection results
- LRSMD detects few spill regions, KIF produces noisy maps, SESC confuses seawater with oil, and R2NMF struggles on some scenes.PCAMD outperforms unsupervised approaches but misses some edge-region oil pixels when trained on 1% of samples.Source: §4.2 · pp. 13, 17
- The proposed method produces the highest detection accuracy overall on HOSD, while PCAMD performs better on five sample images.The visual comparisons cover GM1–GM18, and the proposed method’s spill-region shapes align with reference maps.Source: Figure 4 · p. 14, Figure 5 · p. 15, Figure 6 · p. 16
5. Discussion
The discussion tests severe-noise band removal and ERW optimization. Removing corrupted bands tends to increase detection accuracy, while ERW improves both detection maps and objective scores on GM1.
5.1. The influence of the bands corrupted by severe noise
- Removing severely corrupted bands tends to increase detection accuracy on HOSD.The comparison tests the proposed method with and without those bands.Source: §5.1 · p. 17, Figure 7 · p. 19
5.2. The influence of the ERW optimization
- On GM1, ERW optimization increases AUC by 3.06% and DP by 5.37%.The optimized map also smooths misclassified pixels in homogeneous regions.Source: Figure 8 · p. 19, §5.2 · p. 20
- ERW integrates spatial information so adjacent pixels can belong to the same object in the final detection result.Without ERW, the initial detection map tends to contain salt-and-pepper noise.Source: §5.2 · pp. 17, 20
6. Conclusions
The conclusion presents an unsupervised hyperspectral oil-spill detector and freely available HOSD, reporting performance above several advanced approaches while identifying pseudo-label quality as unresolved.
- The detector removes severely noisy bands, automatically selects training samples without manual labels, and uses adjacent-pixel spatial information to reduce map noise.Source: §6 · p. 20
- The proposed approach outperforms advanced high-dimensional clustering, unsupervised object detection, and supervised oil-spill detection approaches.Source: §6 · p. 20
- The study introduces freely available HOSD, a large-scale oil-spill detection dataset constructed from few images.Source: §6 · p. 20
- Automatically generated training samples are not guaranteed highly accurate, motivating future work on a higher-quality training set.Source: §6 · p. 20