Lightweight 3D Dense Autoencoder Network for Hyperspectral Remote Sensing Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning methods for Hyperspectral Remote Sensing Image (HRSI) classification are hindered by the scarcity of labeled training samples.
• Improving classification accuracy with limited data is crucial for effective HRSI analysis.

Purpose of the Study:

• To propose a novel Lightweight 3D Dense Autoencoder Network (L3DDAN) to enhance HRSI classification accuracy when training samples are limited.
• To develop an effective deep learning framework that overcomes the challenge of insufficient labeled data in HRSI classification.

Main Methods:

• A stacked autoencoder architecture (L3DDAN) combining 3D convolutional operations and 3D dense blocks in the encoder for deep feature extraction.
• A decoder utilizing 3D deconvolution operations for data reconstruction.
• A hybrid training strategy involving successive unsupervised and supervised learning phases, followed by classification using a fine-tuned encoder.

Main Results:

• The proposed L3DDAN framework demonstrated superior performance compared to eight state-of-the-art algorithms on three benchmark HRSI datasets, particularly under limited training sample conditions.
• The network achieved high classification accuracy with significantly fewer trainable parameters than existing methods.
• Successful application of L3DDAN to vegetation classification tasks was shown.

Conclusions:

• The L3DDAN is an effective deep learning model for HRSI classification, especially when labeled training data is scarce.
• The lightweight and efficient design of L3DDAN offers a promising solution for improving hyperspectral image analysis.
• Future research will focus on reducing training time and expanding applications to diverse real-world datasets.