A full convolutional network based on DenseNet for remote sensing scene classification.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Convolutional Neural Networks (CNNs) in remote sensing face challenges with large parameter counts leading to overfitting and insufficient depth for extracting abstract semantic information.
• Existing models often struggle to balance network depth with computational efficiency.

Purpose of the Study:

• To propose a simple, efficient, and deep full convolutional network based on DenseNet for improved remote sensing scene classification.
• To address the limitations of parameter inefficiency and inadequate feature extraction in current CNN models for this domain.

Main Methods:

• A DenseNet-based full convolutional network is designed, utilizing dense connections to generate numerous reusable feature maps with a small number of kernels, enabling greater depth without significant parameter increase.
• An adaptive average 3D pooling operation is incorporated to handle variable input image sizes and reduce feature channels.
• Convolutional layers replace fully connected layers for simplified classification without flattening.

Main Results:

• The proposed network achieves over 100 layers with approximately 7 million parameters, significantly fewer than VGG models.
• The model demonstrates significant improvements in classification performance across the UCM, AID, OPTIMAL-31, and NWPU-RESISC45 datasets compared to state-of-the-art methods.
• The network effectively extracts abstract semantic information and mitigates overfitting issues.

Conclusions:

• The proposed DenseNet-based full convolutional network offers an efficient and effective solution for remote sensing scene classification.
• The method successfully overcomes the limitations of parameter redundancy and insufficient network depth, achieving superior performance.
• The adaptive pooling and convolutional classification further enhance the model's practicality and performance.