UAV Hyperspectral Vegetation Classification via 3D Atrous Vision Transformer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for monitoring grassland degradation using UAV hyperspectral images face challenges in accurately classifying vegetation species due to similar spectral information among species and increased spectral variability with higher spatial resolution, requiring efficient and accurate data processing techniques.
Innovation Solution
The use of a mobile 3D atrous convolution vision Transformer model for UAV hyperspectral image processing, which includes preprocessing, vegetation index fusion, and a multi-level structure design with 3D convolution, atrous convolution, and reverse residual structures to enhance feature extraction and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If the spatial resolution of UAV hyperspectral images is improved, then the detail information of vegetation species is enhanced, but the spectral variability of ground objects increases and intra-class variance of similar grass species intensifies
Solution Approach 1:
The patent transforms the 2D spatial image data into 3D spectral cubes by adding the spectral dimension. This allows the model to capture spectral variations across different wavelengths simultaneously, enabling differentiation of vegetation species based on their unique spectral signatures rather than relying solely on spatial resolution.
Solution Approach 2:
The patent introduces vegetation indices as intermediary features that bridge the gap between raw spectral data and classification outcomes. These indices serve as mediators that highlight key spectral characteristics of different vegetation types, reducing the impact of spectral variability caused by high spatial resolution.
2Measurement precision
If conventional convolutional neural network is used for local feature extraction, then local spatial spectral information is captured, but long-distance dependence of spectral features cannot be established
Solution Approach 1:
The patent merges the strengths of convolutional neural networks (local feature extraction) and vision Transformers (global attention mechanisms) into a hybrid architecture. The CNN component captures local spatial-spectral patterns while the Transformer component establishes long-distance dependencies, allowing the model to benefit from both local and global contextual information.
3Loss of information
If vision Transformer model is used to capture long-distance dependence, then global spectral features are extracted, but local spatial spectral fusion information is not captured well
Solution Approach 1:
The patent segments the feature extraction process into two distinct stages: local feature extraction by CNN and global feature integration by Transformer. This segmentation allows each component to specialize in its strength - CNN handles local spatial-spectral fusion while Transformer captures global dependencies, resolving the contradiction between local and global feature capture.
4Ease of operation
If manual feature design of expert knowledge is used in conventional machine learning, then interpretability is maintained, but accuracy and efficiency in handling high-dimensional nonlinear data structures deteriorate
Solution Approach 1:
The patent employs deep learning models that automatically learn and extract features directly from raw hyperspectral data without requiring manual feature engineering. The model performs self-service by identifying relevant spectral and spatial patterns autonomously, achieving high classification accuracy while handling the complexity of high-dimensional nonlinear data structures.
Data Source
AI summary
The application discloses a classification method and a system of UAV hyperspectral vegetation species based on a deep learning, where the method includes the following steps: collecting hyperspectral images by a UAV; preprocessing collected hyperspectral images to obtain preprocessed images, and performing a stitching mosaicking preprocessing on the preprocessed images to obtain hyperspectral orthoimages; labeling the hyperspectral orthoimages to obtain a label data set; performing a vegetation index fusion on the hyperspectral orthoimages to obtain vegetation index-hyperspectral orthoimages; constructing a grassland vegetation classification model based on the vegetation index-hyperspectral orthoimages and the label data set, and completing a vegetation species classification by using the grassland vegetation classification model.