Hyperspectral Data Compression Using Class-Based Band Decorrelation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for compressing hyperspectral imagery data face challenges due to high computational and processing costs, particularly in decorrelating spectral bands, which are often required for efficient data transmission and storage, and current techniques may lose information or require significant resources for classification.
Innovation Solution
A method that decorrelates hyperspectral data by organizing samples into classes based on values and slopes of preceding components, using residual calculations and adaptive estimation to reduce processing overhead, allowing for efficient compression with reduced side information transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If spectral bands are decorrelated using traditional methods (K-L transform, wavelet transform, or spectrally classified predictors), then compression performance is improved, but computational and processing costs increase significantly
Solution Approach 1:
The patent segments the spectral bands into multiple groups based on their correlation characteristics. Instead of applying complex decorrelation methods to all bands uniformly, the method divides them into segments (groups) where bands within each group have similar correlation properties. This segmentation allows for simpler, group-specific prediction models rather than a single complex global model, thereby reducing overall computational complexity while maintaining compression performance.
Solution Approach 2:
The patent applies local quality by using different prediction strategies for different spectral band groups based on their local correlation characteristics. Each group is assigned a prediction model tailored to its specific properties rather than using a uniform approach across all bands. This localized approach optimizes compression for each group's characteristics while avoiding the computational burden of a single complex global model.
2Loss of information
If spectrally classified predictors are used to exploit nonstationarity of data, then decorrelation performance is improved, but additional passes through the data are required increasing processing time
Solution Approach 1:
The patent performs preliminary classification of spectral bands into groups based on their correlation characteristics before applying prediction. This preliminary action organizes the data structure in advance, allowing subsequent prediction operations to proceed efficiently without requiring multiple passes through the data. The classification is done once, and then prediction is applied directly to each group, eliminating the need for iterative re-processing.
3Device complexity
If simple linear prediction is applied uniformly across the hyperspectral image, then processing complexity is reduced, but compression performance deteriorates due to inability to exploit local nonstationarity
Solution Approach 1:
The patent applies local quality by dividing the spectral bands into groups and applying prediction models tailored to each group's local correlation characteristics. Instead of a single uniform prediction across all bands, each group receives a prediction strategy matched to its specific properties, thereby capturing local nonstationarity without requiring excessively complex global models. This achieves a balance between simplicity and performance.
Solution Approach 2:
The patent segments the spectral bands into multiple groups based on correlation characteristics, allowing simple linear prediction to be applied effectively within each segment. The segmentation enables the simple prediction method to work well locally within groups while the overall system achieves good compression by combining results from multiple groups, thus maintaining simplicity without sacrificing performance.
Data Source
AI summary
A method, apparatus and computer program product are provided for compressing data in a manner that decorrelates a plurality of components of multicomponent data and then encodes data relating to decorrelation of the components. The components may be decorrelated by organizing samples of a component into classes based upon values of samples having corresponding spatial locations in at least one preceding component. In this regard, the organization of samples of a component into classes may include defining classes based upon values of the samples in at least one preceding component. In addition, the method may also define classes based upon the slope of the samples in at least two preceding component.