Hyperspectral Calibration Detection: A Novel Concept For Change Detection With Unsupervised Incremental Safe...

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Hyperspectral change detection (HCD) is crucial for applications like land cover monitoring.
• Current semi-supervised HCD methods require some labeled data, limiting their use in real-time onboard edge computing scenarios.
• Zero-label HCD is needed for immediate detection responses where ground-truth labeling is unavailable.

Purpose of the Study:

• To develop a fully unsupervised and lightweight HCD algorithm for onboard detection missions.
• To address the zero-label requirement in practical HCD scenarios.
• To compensate for variability in acquisition conditions in bitemporal images.

Main Methods:

• Proposed a fully unsupervised HCD algorithm named HyperLUCID (hyperspectral looping unsupervised calibration and incremental detection).
• Employed an iteratively augmented training set to safely collect unchanged pixel samples.
• Learned an iteratively refined spectrum calibration function to compensate for acquisition condition variability.

Main Results:

• HyperLUCID achieved state-of-the-art results with overall accuracy ranging from 93.6% to 97.9% on benchmark HCD datasets.
• The algorithm demonstrated significant computational efficiency, being 1 to 2 orders of magnitude faster than most benchmark HCD methods.
• Successfully compensated for variability in acquisition conditions, enabling easy detection of changed pixels.

Conclusions:

• The proposed HyperLUCID algorithm is highly suitable for onboard detection missions requiring zero-label hyperspectral change detection.
• HyperLUCID offers a computationally efficient and accurate solution for real-time remote sensing applications.
• The algorithm effectively handles variability in bitemporal hyperspectral images, improving change detection performance.