Multitask semantic change detection guided by spatiotemporal semantic interaction.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Semantic Change Detection (SCD) is vital for remote sensing image analysis but faces challenges in capturing spatial details and temporal dependencies.
• Existing methods struggle with change category imbalance and limited accuracy, especially for small targets.
• Inadequate feature extraction and inconsistency between semantic information and change areas hinder performance.

Purpose of the Study:

• To propose a novel network, STGNet, for multitask semantic change detection guided by spatiotemporal semantic interaction.
• To enhance the capture of spatial details and improve feature extraction in complex scenes.
• To resolve inconsistencies between semantic information and change areas for improved detection accuracy.

Main Methods:

• Introduced a Detail-Aware Path (DAP) to enhance spatial detail capture.
• Designed a Bidirectional Guidance Module for adaptive feature selection.
• Developed a Cross-Temporal Refinement Interaction Module (CTIM) with dynamic depthwise separable convolution for cross-time scale feature fusion and interaction.

Main Results:

• STGNet demonstrated superior performance across three SCD datasets compared to existing methods.
• Achieved an F1 score (F1_scd) of 91.64% on the Landsat-SCD dataset.
• Improved the separation Kappa coefficient by 17.68%, indicating enhanced accuracy in detecting changes.

Conclusions:

• STGNet significantly improves semantic change detection accuracy, robustness, and generalization capability.
• The proposed spatiotemporal semantic interaction effectively addresses limitations of previous SCD methods.
• The method shows strong potential for practical applications in remote sensing image analysis.