GETNET: A General End-to-end Two-dimensional CNN Framework for Hyperspectral Image Change Detection (1905.01662v1)

Abstract: Change detection (CD) is an important application of remote sensing, which provides timely change information about large-scale Earth surface. With the emergence of hyperspectral imagery, CD technology has been greatly promoted, as hyperspectral data with the highspectral resolution are capable of detecting finer changes than using the traditional multispectral imagery. Nevertheless, the high dimension of hyperspectral data makes it difficult to implement traditional CD algorithms. Besides, endmember abundance information at subpixel level is often not fully utilized. In order to better handle high dimension problem and explore abundance information, this paper presents a General End-to-end Two-dimensional CNN (GETNET) framework for hyperspectral image change detection (HSI-CD). The main contributions of this work are threefold: 1) Mixed-affinity matrix that integrates subpixel representation is introduced to mine more cross-channel gradient features and fuse multi-source information; 2) 2-D CNN is designed to learn the discriminative features effectively from multi-source data at a higher level and enhance the generalization ability of the proposed CD algorithm; 3) A new HSI-CD data set is designed for the objective comparison of different methods. Experimental results on real hyperspectral data sets demonstrate the proposed method outperforms most of the state-of-the-arts.

• The paper presents the GETNET framework, which integrates subpixel information through a mixed-affinity matrix to enhance change detection in hyperspectral imaging.
• It employs a novel 2D CNN for multi-source feature extraction, achieving a 3% to 10% overall accuracy improvement over traditional methods.
• The curation of a new 'river' dataset enables robust benchmarking, underscoring the framework’s practical applications in remote sensing and land cover analysis.

An Analysis of GETNET: A General End-to-end Framework for Hyperspectral Image Change Detection

The paper "GETNET: A General End-to-end Two-dimensional CNN Framework for Hyperspectral Image Change Detection" introduces a novel approach to the challenging task of detecting changes in hyperspectral imaging (HSI). This is a critical topic in remote sensing due to the complexity and high dimensionality of hyperspectral data, which pose significant challenges for conventional change detection (CD) methodologies. This paper presents the GETNET framework, which leverages deep learning to process multi-source information for superior performance in change detection tasks.

Key Contributions

The authors conceptualize and develop a comprehensive end-to-end deep learning framework aimed at improving hyperspectral image change detection (HSI-CD). The major contributions of the paper can be classified into three significant advancements:

1. Mixed-Affinity Matrix: This framework introduces a mixed-affinity matrix that ambitiously integrates subpixel information to capture gradient features by converting the hyperspectral data and abundance maps to two-dimensional matrices. These matrices harness cross-channel gradient information more effectively, addressing both pixel and subpixel level representations. This approach departs from traditional 1-D vector analysis, providing richer representational capability for CD tasks.
2. 2-D CNN for Multi-source Feature Learning: A two-dimensional Convolutional Neural Network (CNN) is crafted for feature extraction from the complex hyperspectral data. This method efficiently mines discriminative features while enhancing the model's generalization ability. It extends the capability of existing techniques by learning significant spectral patterns through convolution operations tailored for the mixed-affinity matrices.
3. Curation of New Data Set: The authors contribute to the field by designing a new hyperspectral data set, titled "river", to facilitate objective benchmarking of various HSI-CD methods. This data set provides a complex tapestry of changes, ensuring robust testing of change detection algorithms.

Experimental Validation and Performance Insights

The GETNET framework demonstrates significant improvement over several established methods such as CVA, PCA-CVA, IR-MAD, and even generic CNN models.Results on real hyperspectral data sets reveal GETNET’s capability to outperform these state-of-the-art approaches consistently. A notable observation is the framework’s superior overall accuracy (OA) and Kappa coefficient, indicating robust performance in both detecting changed and unchanged regions across diverse data sets:

• Across the "farmland", "countryside", and "Poyang lake" data sets, GETNET achieves a remarkable OA consistently outperforming conventional methods by approximately 3% to 10%, depending on the data set complexity.
• In the ablation paper, the inclusion of subpixel information through unmixing significantly boosts the framework’s accuracy, showcasing the importance of detailed spectral analysis.

Implications and Future Directions

GETNET's architecture provides a sophisticated mechanism for handling high-dimensional hyperspectral data, suggesting promising applications in various geospatial fields such as disaster monitoring, resource exploration, and land cover change detection. The framework aligns with the increasing trends of integrating deep neural network models with rich data representations for comprehensive task performance improvement.

Future research directions could explore the incorporation of additional context-aware models that leverage temporal dynamics in hyperspectral data. Moreover, expanding the framework's flexibility and adaptability to real-time applications with on-the-fly training capabilities could enhance utility.

Conclusion

The GETNET framework emerges as a competent solution for hyperspectral image change detection, pushing the boundaries of existing approaches through innovative use of mixed-affinity matrices and deep learning techniques. While the paper does not sensationalize its contributions, the presented advancements mark substantial progress in hyperspectral data analysis, laying groundwork for future refinements and applications in remote sensing.