Adaptive Multi-Scale Feature Extraction for Hyperspectral Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Hyperspectral image classification faces challenges due to varying spatial resolutions and scale variability of land covers, with existing methods failing to adaptively extract multi-scale features across different datasets, leading to limited classification accuracy.
Innovation Solution
An adaptive multi-scale feature extraction model is introduced, utilizing a scale reference network and two feature extraction networks with conditional gating mechanisms and convolutional pooling modules, allowing for automatic adjustment of the network structure to extract features at different scales, and employing the RMSProp algorithm for loss minimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a fixed network model is used for feature extraction from different hyperspectral datasets, then the model structure is simple and easy to implement, but the classification accuracy deteriorates due to inability to adapt to heterogeneous spatial resolutions
Solution Approach 1:
The patent implements dynamic adaptability by introducing a scale reference network that automatically adjusts the network structure based on the spatial resolution of input hyperspectral datasets. The conditional gating mechanism dynamically selects and fuses features from multiple convolutional pooling modules with different kernel sizes, allowing the model to adapt to heterogeneous spatial resolutions rather than using a fixed network structure.
Solution Approach 2:
The patent changes structural parameters adaptively by using the scale reference network to determine which convolutional pooling modules to activate based on the input data's spatial resolution characteristics. The conditional gating mechanism adjusts feature fusion parameters dynamically, changing the network's effective architecture to match the input dataset's properties.
2Measurement precision
If multi-scale feature extraction is implemented to handle scale variability of land covers, then the classification accuracy improves, but the device complexity increases due to multiple networks and adaptive mechanisms
Solution Approach 1:
The patent segments the feature extraction process into multiple parallel convolutional pooling modules with different kernel sizes (3x3, 5x5, 7x7), each capturing features at different scales. The scale reference network segments the decision-making process by evaluating spatial resolution characteristics and selectively activating appropriate modules, making the complex multi-scale extraction manageable and adaptive.
Solution Approach 2:
The patent implements a nested structure where the scale reference network is embedded within the overall feature extraction framework, and the conditional gating mechanism is nested within each convolutional pooling module. This nested architecture allows the system to manage complexity by organizing multiple scales and decision layers in a hierarchical manner.
3Ease of manufacture
If uniformly designed network modules are used for all datasets, then the model design is straightforward, but information extraction accuracy deteriorates due to ignoring spatial resolution differences
Solution Approach 1:
The patent applies local quality by designing different convolutional pooling modules with specific kernel sizes (3x3, 5x5, 7x7) optimized for different spatial resolution characteristics. The scale reference network evaluates local properties of the input dataset and activates appropriate modules, ensuring that each part of the network is optimized for its specific function and the input data's characteristics.
Data Source
AI summary
Disclosed is a method for classifying hyperspectral images on the basis of an adaptive multi-scale feature extraction model, the method comprising: establishing a framework comprising the two parts of a scale reference network and a feature extraction network, introducing a condition gate mechanism into the scale reference network, performing determination step-by-step by means of three groups of modules, inputting features into a corresponding scale extraction network, deep mining rich information contained in a hyperspectral remote sensing image, effectively combining features of different scales, improving a classification effect, and generating a fine classification result map.