Exploitation of Class Activation Map to Improve Land Cover and Land Use Classification Using Deep Learning
This study investigates the potential of gradient-weighted class activation mapping (Grad-CAM++) in enhancing land cover and land use (LCLU) classification using deep learning models. A U-Net and an Attention U-Net model were trained on Sentinel-2 imagery to classify 10 LCLU classes in a study area...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2024-12-01
|
| Series: | Proceedings |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2504-3900/110/1/3 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This study investigates the potential of gradient-weighted class activation mapping (Grad-CAM++) in enhancing land cover and land use (LCLU) classification using deep learning models. A U-Net and an Attention U-Net model were trained on Sentinel-2 imagery to classify 10 LCLU classes in a study area in Northern Ontario, Canada (centered at 49.17° N, 83.03° W). The classes included water, wetland, deciduous forest, mixed forest, coniferous forest, barren, urban/development, agriculture, shrubland, and no data (masked areas). The U-Net model achieved overall accuracy of 70.68%, a mean intersection over union (IoU) of 0.4852, and an F1 score of 0.7150, slightly outperforming the Attention U-Net model. Grad-CAM++ visualizations revealed that both models correctly focused on relevant features for each LCLU class, enhancing the interpretability of deep learning models in remote sensing applications. The findings suggest that integrating Grad-CAM++ with deep learning architectures can improve model transparency and guide future enhancements in LCLU classification tasks. |
|---|---|
| ISSN: | 2504-3900 |