Deep-Growing Neural Network With Manifold Constraints for Hyperspectral Image Classification.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Deep neural networks (DNNs) struggle with overfitting due to insufficient labeled data.
• Semisupervised learning methods utilize unlabeled data to mitigate label scarcity.
• Traditional models face challenges adapting their fixed structure to growing pseudolabel sets.

Purpose of the Study:

• To propose a novel deep-growing neural network with manifold constraints (DGNN-MC) for semisupervised learning.
• To enable networks to dynamically adjust depth based on available pseudolabels.
• To preserve local structure in high-dimensional data during semisupervised learning.

Main Methods:

• A framework filters shallow network outputs to generate high-confidence pseudolabeled samples.
• The network depth increases iteratively as the pseudolabeled training set grows.
• Manifold constraints are applied to preserve data structure during network growth.

Main Results:

• The proposed DGNN-MC dynamically deepens network structure to match increasing pseudolabel pools.
• Experimental results on Hyperspectral Image (HSI) classification demonstrate superior performance.
• The method effectively balances network learning capacity with growing labeled data.

Conclusions:

• The DGNN-MC offers an effective solution for semisupervised learning challenges.
• The dynamic network growth approach enhances model adaptability and performance.
• This method shows significant potential for applications requiring efficient utilization of limited labeled data.