Ship extraction at sea based on hyperspectral satellite remote sensing

Hyperspectral image classification methods differ fundamentally from traditional multispectral classification. A continuous spectral curve can be obtained from each pixel of a hyperspectral image. This allows for comparison between known spectral curves and the spectral curves obtained from each pixel in the image. Ideally, if the two spectral curves are similar, it can indicate which substance the pixel is closer to.

The characteristics of hyperspectral images allow them to be applied not only to general image classification but also to material identification and target detection. Image classification focuses more on ground cover and material composition, while target identification and detection search for specific objects, with the result being "present" or "absent." Therefore, we refer to hyperspectral image classification, material identification, and detection as spectral recognition.

In maritime vessel identification scenarios, hyperspectral spectral identification offers a promising approach to identify maritime targets and distinguish different types of vessels in complex maritime environments.

Specific procedures

1. Atmospheric correction

Use one of the following methods for atmospheric correction:

Quick Atmospheric Correction (QUAC)

Radiative transfer models (6S, FLAASH)

Statistical atmospheric correction methods (IAR Reflectance, Log Residuals, Flat Field, Empirical Line)

Simplified Dark Subtraction

2. Select target samples, calculate target and background spectra.

When applying Orthogonal Subspace Projection (OSP), Target-Constrained Interference-Minimized Filter (TCIMF), and Mixture Tuned Target-Constrained Interference-Minimized Filter (MTTCIMF), at least two target spectra are required, or additionally, spectra that are easily confused with the target spectra should be selected as background spectra to help improve detection accuracy.

3. MNF Transformation

MNF transformation can separate noise and reduce the dimensionality of data to reduce computation. The Mixture Tuned Matched Filter (MTMF) and Mixture Tuned Target-Constrained Interference-Minimized Filter (MTTCIMF) identification methods are based on MNF.

Specific procedures

4. Choose appropriate analysis methods

Common methods are as follows:

Matched Filtering (MF)

Constrained Energy Minimization (CEM)

Adaptive Coherence Estimator (ACE)

Spectral Angle Mapper (SAM)

Orthogonal Subspace Projection (OSP)

Target-Constrained Interference-Minimized Filter (TCIMF)

Mixture Tuned TCIMF (MTTCIMF)

Mixture Tuned Matched Filtering (MTMF)

5. Target Extraction

By using a rule-based threshold or by selecting point clouds with high MF scores and low infeasibility values ​​from a scatter plot based on MF scores and infeasibility values, the detected targets can be identified.

6. Post-processing

Small blobs in the results are removed using post-classification processing methods (convolutional and qualitative methods, etc.). Simultaneously, the minimum number of cluster pixels is set based on the target size to remove isolated small blobs. Vectorization and other transformations are performed as needed.

XIOPM SPACE

Extraction results

▲True-color composite of original hyperspectral images

▲ Target Areas 1 and 2

▲ Target areas 3 and 4

▲ Target Area 5

▲ Spectral curves of each target and background

Referring to the image above, the spectral curves of target 2 and target 4 are relatively similar, suggesting they are of the same type of vessel. While hyperspectral spectral identification holds great promise, several challenges remain in identifying vessels at sea.

1. Data volume and processing: The large amount of data generated by hyperspectral sensors requires efficient processing algorithms and computing resources.
2. Spectral variability: Due to the influence of factors such as material composition and sea surface roughness, spectral characteristics may vary, thus requiring robust classification algorithms.
3. Data availability: For maritime applications, acquiring high-quality hyperspectral images is crucial for training and validating classification models.
4. Data fusion: Hyperspectral data needs to be fused with other information sources (such as high-resolution optical and radar data) to improve recognition accuracy.