DeepEarthMY: A Remote Sensing Dataset for Tropical Land-Cover Segmentation
Land-cover mapping is essential for applications such as urban planning, natural resource management, and environmental monitoring. However, tropical equatorial regions pose unique challenges to land-cover classification due to dense vegetation, persistent cloud cover, and spectral similarity across...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10946984/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Land-cover mapping is essential for applications such as urban planning, natural resource management, and environmental monitoring. However, tropical equatorial regions pose unique challenges to land-cover classification due to dense vegetation, persistent cloud cover, and spectral similarity across land-cover types. Moreover, existing land-cover classification models often perform poorly in these regions as they are primarily trained on datasets from non-equatorial climates. To address this gap, we introduce DeepEarthMY, a high-resolution land-cover mapping dataset specifically designed for equatorial regions. The dataset contains 4,007 meticulously annotated images from 52 diverse regions across Malaysia, representing key land-cover types such as forests, buildings, roads, water bodies, agricultural lands, and barren land. We evaluated the dataset by performing land-cover segmentation on selected state-of-the-art semantic segmentation models, including DC-Swin, HRNet, and UNetFormer. The experimental results reveal that the DC-Swin model achieves the best performance, with a mIoU score of 80.23%. To further evaluate the significance of region-specific datasets for land-cover classification, we performed cross-dataset testing on two datasets: DeepEarthMY from the equatorial region and LoveDA from the temperate climate, using five-fold cross-validation. HRNet models trained on DeepEarthMY achieved a mIoU of 77.2% on DeepEarthMY but only 33.2% on the LoveDA test set. Conversely, models trained on LoveDA achieved 51.4% and 43% on the LoveDA and DeepEarthMY test sets, respectively. This significant performance gap highlights the need for region-specific datasets to advance land-cover research, particularly in equatorial climates where land-cover datasets are scarce. In conclusion, the DeepEarthMY dataset can be an invaluable resource for remote sensing in the equatorial region. The dataset is available at <uri>https://zenodo.org/records/14242124</uri>. |
|---|---|
| ISSN: | 2169-3536 |