Hyperspectral Image Classification with Attention Aided CNNs (2005.11977v2)

Abstract: Convolutional neural networks (CNNs) have been widely used for hyperspectral image classification. As a common process, small cubes are firstly cropped from the hyperspectral image and then fed into CNNs to extract spectral and spatial features. It is well known that different spectral bands and spatial positions in the cubes have different discriminative abilities. If fully explored, this prior information will help improve the learning capacity of CNNs. Along this direction, we propose an attention aided CNN model for spectral-spatial classification of hyperspectral images. Specifically, a spectral attention sub-network and a spatial attention sub-network are proposed for spectral and spatial classification, respectively. Both of them are based on the traditional CNN model, and incorporate attention modules to aid networks focus on more discriminative channels or positions. In the final classification phase, the spectral classification result and the spatial classification result are combined together via an adaptively weighted summation method. To evaluate the effectiveness of the proposed model, we conduct experiments on three standard hyperspectral datasets. The experimental results show that the proposed model can achieve superior performance compared to several state-of-the-art CNN-related models.

• The paper introduces a dual-branch CNN that incorporates spectral and spatial attention mechanisms to enhance hyperspectral image classification performance.
• It employs lightweight attention modules using 1-D and 2-D convolutions to selectively refine crucial spectral channels and spatial features.
• Experimental results on standard datasets show improved overall accuracy and robustness against spectral variability and spatial heterogeneity.

Hyperspectral Image Classification with Attention Aided CNNs: A Summary

The paper "Hyperspectral Image Classification with Attention Aided CNNs" presents a sophisticated approach to hyperspectral image classification involving convolutional neural networks (CNNs) enhanced with attention mechanisms. The central contribution is the development of a model that integrates spectral and spatial attention modules to improve classification accuracy by focusing on the most discriminative spectral bands and spatial positions.

Methodological Overview and Key Components

1. Spectral and Spatial Attention Mechanisms: The model implements two types of attention modules:
   • Spectral Attention: This module sharpens the focus on the spectral domains, enhancing features from important channels and suppressing irrelevant ones. It leverages global pooling layers and 1-D convolutions to achieve spectral feature refinement.
   • Spatial Attention: In parallel, the spatial attention module targets spatial feature enhancement. Using 1×1 convolution and subsequent 2-D convolutions, it refines the spatial focus within feature maps.
2. Two-Branch CNN Framework: By constructing separate spectral and spatial attention sub-networks, the model efficiently captures and processes both spectral and spatial information. Each sub-network independently classifies the hyperspectral data, and the results are combined through a weighted summation, allowing adaptive integration of spectral and spatial classification performance.
3. Lightweight Design and Optimization: The proposed attention modules are designed to be computationally lightweight, making them suitable for hyperspectral datasets with limited training samples. They are integrated with relatively shallow CNNs to mitigate overfitting, often encountered in deep networks with limited data.

Experimental Evaluation and Results

The paper details three extensive experiments conducted on standard datasets: Houston 2013, Houston 2018, and HyRANK, with findings suggesting marked improvement in performance metrics like overall accuracy (OA), average accuracy (AA), and Kappa coefficient compared to existing models. Notably, the integration of attention mechanisms enabled the CNNs to outperform conventional spectral-spatial models and refine feature extraction capability, which alleviated common issues like spectral variability and spatial heterogeneity.

Implications and Future Directions

This research contributes a robust framework for hyperspectral image analysis, illustrating how attention mechanisms can enhance CNN performance by effectively modeling the importance of different spectral and spatial data aspects. Practically, such advances can lead to more accurate land cover classification, impacting fields such as environmental monitoring and agricultural management. Theoretically, the paper suggests directions for further refinements in attention modules' design and integration into deeper networks. Future research could explore hybrid architectures combining attention-aided CNNs with recurrent models for temporal hyperspectral data or investigate more automated ways of determining weighted combinations of spectral and spatial results. As computational capacities and available training sets grow, extending these methods to deeper models and larger datasets remains a promising avenue.