Double-Branch Multi-Scale Dual-Attention Network for Lithology Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for surface lithology identification in hyperspectral remote sensing images suffer from low identification accuracy due to the use of simple convolutional networks, which struggle to extract complex high-dimensional spectral and spatial features, especially with limited sample sizes and potential information leakage.
Innovation Solution
An object-oriented method using a double-branch multi-scale dual-attention mechanism network (DBMSDA) is employed to extract and fuse spectral and spatial features, comprising a spectral branch with multi-scale spectral residual attention and a spatial branch with spatial attention mechanisms, ensuring no information leakage through a dataset division strategy that separates training and test sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If simple convolutional networks are used for deep learning, then the model structure is simple and easy to implement, but the identification accuracy of spectral and spatial features is low due to inability to extract complex high-dimensional features
Solution Approach 1:
The network is segmented into two independent branches: a spectral branch for extracting spectral features and a spatial branch for extracting spatial features. Each branch processes specific feature types through dedicated convolutional layers, allowing complex high-dimensional feature extraction while maintaining manageable model complexity through modular architecture.
Solution Approach 2:
The patent transitions from traditional single-dimension feature extraction to multi-dimensional feature fusion. The spectral branch extracts spectral dimension features, the spatial branch extracts spatial dimension features, and the fusion layer combines these dimensions, enabling the model to capture complex high-dimensional spectral-spatial relationships that simple networks cannot handle.
2Reliability
If manual on-site exploration and inspection is used for lithology identification, then the method is simple and reliable, but the time and material costs are high
Solution Approach 1:
The patent replaces manual on-site exploration with an automated deep learning system. The double-branch multi-scale dual-attention mechanism network processes hyperspectral remote sensing images automatically, substituting human expertise with computational models that can rapidly analyze spectral and spatial features without physical fieldwork.
Solution Approach 2:
The system creates a digital copy of the physical inspection process through the neural network model. The model learns from training data representing various lithology types and reproduces identification capabilities, allowing rapid classification of new images without requiring physical presence or manual verification for each sample.
3Productivity
If machine learning algorithms like PCA, MNF, and ICA are used, then manual operations are reduced and efficiency is improved, but the identification accuracy remains limited compared to deep learning approaches
Solution Approach 1:
The patent changes the fundamental parameters of the learning approach by transitioning from traditional machine learning algorithms to deep learning with custom architecture. The double-branch multi-scale dual-attention mechanism introduces new parameters including multi-scale convolution kernels, residual connections, and attention weights, enabling the model to capture complex non-linear relationships in hyperspectral data that traditional algorithms cannot handle.
Data Source
AI summary
The present disclosure provides an object-oriented method for identifying and classifying surface lithology in a hyperspectral remote sensing image. The method includes: determining a hyperspectral remote sensing image in a research area and a lithology type label corresponding to each hyperspectral remote sensing image, and preparing a hyperspectral remote sensing dataset; dividing pixels of the hyperspectral remote sensing image in the hyperspectral remote sensing dataset into a training set and a test set through a division strategy for a dataset without leakage information; and based on a deep learning method, extracting and fusing, through a double-branch multi-scale dual-attention mechanism network based on the training set and the test set, a spectral feature and a spatial feature of a hyperspectral remote sensing image to be tested, to generate a fused feature for representing a surface lithology type of the hyperspectral remote sensing image to be tested.