Research on Ground Object Classification Method of High Resolution Remote-Sensing Images Based on Improved DeeplabV3

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• High-resolution remote-sensing image segmentation is crucial for land planning and ecological monitoring.
• Traditional methods struggle with complex scenes, and existing deep learning models like DeeplabV3+ have limitations in accuracy and training cost.
• Improving semantic segmentation for remote sensing is an active research area.

Purpose of the Study:

• To enhance the DeeplabV3+ network for improved semantic segmentation of high-resolution remote-sensing images.
• To address the limitations of poor segmentation accuracy and high training costs associated with the traditional DeeplabV3+ network.
• To develop a more efficient and accurate deep learning model for ground-object classification in remote sensing.

Main Methods:

• The study improves the DeeplabV3+ network by using MobileNetV2 as the backbone feature-extraction network.
• An attention-mechanism module is integrated after the feature-extraction and ASPP modules.
• Focal loss is introduced to balance the network's learning process.

Main Results:

• The proposed method demonstrates enhanced feature extraction capabilities.
• Network training costs are significantly reduced compared to the traditional DeeplabV3+.
• Experiments show improved mean Intersection over Union (mIoU) scores on benchmark datasets (WHDLD and CCF BDCI), outperforming traditional DeeplabV3+, U-NET, PSP-NET, and MACU-net.

Conclusions:

• The enhanced DeeplabV3+ network offers superior semantic segmentation accuracy for high-resolution remote-sensing images.
• The integration of MobileNetV2 and attention mechanisms, along with focal loss, effectively reduces computational overhead and training time.
• This improved model provides a more efficient and accurate solution for ground-object classification in critical applications like land planning and environmental monitoring.