[NS][Lab_Seminar_240930]ACGT-Net: Adaptive Cuckoo Refinement-Based Graph Transfer Network for Hyperspectral Image Classification.pptx

ACGT-Net: Adaptive Cuckoo Refinement-Based Graph Transfer Network for Hyperspectral Image Classification

1. ACGT-Net: Adaptive Cuckoo Refinement-Based Graph Transfer Network for Hyperspectral Image Classification Tien-Bach-Thanh Do Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea E-mail: os fa19730 @catholic.ac.kr 202 4/09/30 Yuanchao Su et al. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 2023

2. Introduction Hyperspectral Image Classification (HIC): Hyperspectral images capture detailed spectral information for each pixel across multiple bands Widely used in land use/cover research, agriculture, and environmental monitoring Challenges High-dimensional data leading to redundancy Traditional ML methods struggle with generalization and noise GNNs offer a new approach but are limited by rigid and noisy graph structures

3. Goal & Solution Goal Develop a robust graph-based model for hyperspectral image classification Improve graph structure through meta-heuristic optimization Solution ACGT-Net: A Graph Transfer Network enhanced with Adaptive Cuckoo Refinement (ACR) for graph structure refinement

4. Overview Graph Convolutional Network (GCN): Pretrained to learn initial transferable weight parameters Cuckoo Search Strategy (CSS): refine the graph structure to reduce noise and focus on significant features Transfer Learning: leverage knowledge from pretrained GCN to improve generalization on new datasets

5. Model Fig. 1. Flowchart of the proposed ACGT-Net.

6. Model Upstream task: Pretraining a GCN using spatially weighted data Output primary features from the source GCN Downstream task: Refined using CSS-based graph structure refinement Combine primary and intermediate features to create a final multimodal representation Final output: classified hyperspectral image pixels using a softmax layer

7. Graph Construction Node: pixels in the hyperspectral image Edge: define similarities between pixel based on spectral signatures Graph Structure: Built using Spatial Weighted Mean Filter (SWMF) to capture local spatial correlation Let X = [X1, X2, …, Xd] be an HSI, where xi denote pixel and Xi represent a band, before implement the SWMF, normalize nodes to [0, 1] Apply SWMF Adjacency matrix calculated with Radial Basis Function (RBF) where is width of a Gaussian kernel Adopt a batchwise scheme to train

8. Transfer Learning Fig. 2. Illustrations of weight updates and information propagations. The source GCN updates all parameters and conducts forward and backward propagations, while the target GCN only updates weight parameters and implements backward propagation in the (g + 1)th layer.

9. Transfer Learning Pretraining: a source GCN is trained on the initial graph to extract primary features Knowledge Transfer: the source GCN’s learned parameters are transferred to the target GCN Hybrid Domain Adaptation: Instance-based: samples are adjusted between source and target domains Feature-based: primary features are passed from source GCN to target GCN Parameter-based: source GCN provide initial weights to the target GCN Assuming B is degree matrix Latent features

10. Transfer Learning Output features of the g-th GCN layer Loss function number of training sample sample vector one-hot label vector probability of node belong to class Instance-based procedure

11. Graph Refinement using Cuckoo Search Strategy (CSS) Fig. 3. Illustration of CSS-based GSR. The CSS can enable the target GCN to pay more attentions to the significant features. The gray squares represent 1, and the goldenrod squares represent 0. After completing the CSS, we only reserve the latent feature components corresponded to marks with 1.

12. Graph Refinement using Cuckoo Search Strategy (CSS) CSS overview Inspired by cuckoo bird’s parasitic nesting behavior Adjust graph structure by removing noisy edges and refining important ones Process Nodes and edges are refined iteratively using Levy flight (random walk) Optimized features are marked based on their importance (1 or 0) Logistic function used to assign significance scores to features Outcome: a more robust graph structure with reduced noise and enhanced focus on important channels

13. Graph Refinement using Cuckoo Search Strategy (CSS) Let l represent a generation of CSS Step size Location host nest Loss function

14. Experiments Experimental Settings Datasets: Indian Pines Pavia University Xiong’an Matiwan Houston University Metrics: Overall Accuracy Average Accuracy Kappa Coefficient

15. Experiments Results

16. Experiments Results

17. Experiments Results

18. Conclusion Key Contributions: Introduce meta-heuristic optimization into GCN-based HIC Adaptively refined graph structures, enhancing classification performance Demonstrated improved generalization using transfer learning Future Work: Investigating reinforcement learning with meta-heuristic strategies for further enhancement