Context-Rich Network for Hyperspectral Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Hyperspectral image classification methods struggle to effectively capture and utilize context information across different spatial locations and scales, limiting the accuracy of classification due to the use of shallow features and local operators in traditional networks.
Innovation Solution
A context-rich network method that includes a multi-scale feature extraction module for intra-feature and inter-feature relationship capturing, using non-local self-attention mechanisms and a scale context-aware module to extract spatial and scale context information, followed by a classification module for end-to-end hyperspectral image classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional classification methods directly input original spectral vectors into the classifier, then the method is simple and fast, but the classification accuracy is limited due to lack of surrounding environment information
Solution Approach 1:
The patent embeds multiple processing stages within a unified network architecture. The feature extraction module contains convolutional layers that process spectral data, which is then fed into attention mechanism modules that process spatial relationships, creating a nested structure where each module handles a specific aspect of feature processing while contributing to the overall classification accuracy.
Solution Approach 2:
The patent transitions from processing only spectral data (1D) to processing both spectral and spatial information (2D). By introducing spatial attention mechanisms and multi-scale feature extraction, the model adds a spatial dimension to the classification process, enabling it to capture contextual information from surrounding pixels while maintaining spectral discrimination capability.
2Measurement precision
If feature engineering methods extract spatial and spectral features simultaneously, then more information is captured, but the extracted features are shallow and not stable under complex conditions
Solution Approach 1:
The patent replaces traditional hand-crafted feature engineering with a data-driven deep learning approach. Instead of manually designing feature extraction algorithms, the model uses convolutional neural networks to automatically learn hierarchical feature representations from raw spectral data, achieving both deep feature extraction and robustness to complex conditions through end-to-end training.
Solution Approach 2:
The attention mechanism modules automatically identify and emphasize important spatial regions and spectral bands without human intervention. The self-attention mechanism dynamically adjusts feature weights based on the input data characteristics, enabling the model to adapt to complex conditions and maintain reliable feature representations across different scenarios.
3Ease of operation
If convolutional neural networks use spatial patches centered on pixels to be classified, then local perception is achieved, but patches of adjacent pixels have overlapping areas bringing unnecessary and redundant calculations
Solution Approach 1:
The patent extracts and processes only the unique spectral information at each pixel location through 1D convolution, avoiding the redundant computation of overlapping spatial patches. By separating spectral feature extraction from spatial context analysis, the model eliminates redundant calculations while maintaining local perception capability through targeted spatial attention mechanisms.
4Area of stationary object
If convolutional neural networks increase receptive field by stacking convolutional layers, then more spatial context is captured, but long-range dependencies and non-local context information cannot be effectively modeled
Solution Approach 1:
The patent introduces a self-attention mechanism that operates in a different dimensional space compared to traditional convolutional stacking. Instead of expanding receptive field through spatial convolution only, the attention mechanism creates a global context representation by computing relationships between all pixels simultaneously, effectively capturing long-range dependencies and non-local context information through a different computational dimension.
Data Source
AI summary
Disclosed in the present invention is hyperspectral image classification method based on context-rich networks. The method comprises a training stage and a prediction stage, wherein the training stage comprises image pre-processing, sample selection and network training. Firstly, performing normalization on a hyperspectral image, and then randomly selecting an appropriate proportion of marked samples from each category to generate a label map, and performing training by using the designed network; in the prediction stage, directly inputting the whole image into the trained network and obtaining a final classification result. By means of the present invention, data pre-processing, feature extraction, the process of context-rich information capturing, and classification are taken into comprehensive consideration in the whole flow; and the classification of a hyperspectral image is realized by means of constructing an end-to-end network.