Multitask Change-Aware Network and Semisupervised Enhanced Multistep Training for Semantic Change Detection
Semantic change detection (SCD) aims to find out where and what changes between a pair of co-registered remote sensing images. Compared to binary change detection, which only predicts the location of changes, SCD provides detailed from-to change information, helping to gain a comprehensive understan...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10938184/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Semantic change detection (SCD) aims to find out where and what changes between a pair of co-registered remote sensing images. Compared to binary change detection, which only predicts the location of changes, SCD provides detailed from-to change information, helping to gain a comprehensive understanding and analysis of land cover and land use. SCD is a challenging task due to the complexity of scenes in remote sensing images and the lack of semantic labels in SCD datasets. In this work, we propose a model named Multitask Change-Aware Network (MTCAN) and a Multistep Training (MST) method for land cover semantic change detection in optical remote sensing images. To better identify fine-grained semantic changes, the MTCAN comprises feature aggregation module (FAM), spatial enhancement module (SEM), and change extraction module (CEM). FAM integrates low-level spatial details and high-level semantics from multilevel features, which helps to capture small-sized changes. SEM models long-range correlations and global context, providing global representations in binary change detection and semantic segmentation branches. CEM extracts discriminative change features by calibrating differential features with channel and spatial attention, which helps to accurately locate change areas. MST is designed to overcome the insufficient training caused by the lack of semantic labels, consisting of contrastive loss and iterative self-training. The contrastive loss supervises the semantic segmentation parts with binary change labels. In the self-training process, the trained student model is added to the teacher model ensemble that generates pseudo labels for unlabeled areas, which are then used to train the next student. MTCAN-MST achieves 23.48% SeK on SECOND dataset and 67.74% SeK on Landsat-SCD dataset, outperforming the state-of-the-art methods with lower computational cost. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |