A Semantic Change Detection Network Based on Boundary Detection and Task Interaction for High-Resolution Remote...

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Current semantic change detection (SCD) methods often separate semantic segmentation (SS) and change detection (CD) tasks.
• This separation overlooks the potential benefits of task correlation in improving model performance.
• Existing approaches may not fully leverage the interplay between identifying change locations and classifying land cover types.

Purpose of the Study:

• To propose a novel semantic change detection network (BT-SCD) that enhances task correlation using multitask learning.
• To improve the accuracy and robustness of land cover and land use change detection.
• To address the limitations of mainstream SCD approaches by incorporating boundary detection.

Main Methods:

• Introduced a boundary detection (BD) task to strengthen the relationship between SS and CD tasks within the SCD framework.
• Proposed pixel-level and logit-level interaction strategies to facilitate information exchange between SS and CD tasks.
• Developed a bidirectional change feature extraction module to capture temporal change information effectively.

Main Results:

• The proposed BT-SCD network achieved state-of-the-art performance on three standard datasets and a specialized NAFZ dataset.
• Integration of BD task and interaction strategies demonstrated positive reinforcement between SS and CD tasks.
• The bidirectional feature extraction module effectively handled bitemporal features.

Conclusions:

• The BT-SCD model offers a significant advancement in semantic change detection by leveraging multitask learning and task interaction.
• The proposed methods effectively enhance the understanding of land cover and land use dynamics.
• BT-SCD provides a more robust and accurate solution for complex geospatial change analysis.