This repository implements DeepLabV3+ for high-resolution land cover mapping using the OpenEarthMap dataset, which consists of 5,000 aerial and satellite images with 8-class annotations across 97 regions from 44 countries. Models trained on OpenEarthMap generalize globally, making them applicable to urban planning, land cover classification, and environmental monitoring.
- OpenEarthMap Dataset
A comprehensive benchmark dataset featuring 5,000 aerial and satellite images with 8-class manually annotated land cover labels. Covering 97 regions across 44 countries and 6 continents, it offers a global perspective on land cover classification, ensuring robust generalization capabilities for models trained on it.
-
Custom Datasets
-
Scaled Down OpenEarthMap A smaller version of the original dataset has been preprocessed, retaining all details but resized and converted into
jpgandpngformats for efficient training. The processed dataset is available HuggingFace. -
City-Specific Dataset Over 10,000 remote sensing images annotated across six cities for enhanced segmentation tasks. This dataset can be found HuggingFace.
-
We utilized the Scaled Down OpenEarthMap dataset for training, which is a smaller, preprocessed version of the original OpenEarthMap dataset, resized and converted into jpg and png formats for efficient training.
For inference, we applied our trained model to the City-Specific Dataset, which contains over 10,000 remote sensing images annotated across six cities, enhancing segmentation tasks.
The repository contains two primary modules:
Implements the DeepLabV3+ training pipeline using the DeeplabV3Plus_RS_Train.ipynb notebook with the following features:
- Pre-trained weights:
deeplabv3_finetuned_RS_openearthmap_v2.pth,are not included in the GitHub repository due to their large size. You can download the file from the Google Drive: Download Pre-trained Weights - Configuration for semantic segmentation with 8 land cover classes
The DeeplabV3Plus_RS_Predict.ipynb notebook is used for batch processing of satellite images, generating land cover segmentation predictions in .png format with color-coded classes.
The model generates segmentation masks with the following classes:
- Background: Black (RGB: 0, 0, 0)
- Bareland: Dark Red (RGB: 128, 0, 0)
- Rangeland: Bright Green (RGB: 0, 255, 36)
- Developed Space: Gray (RGB: 148, 148, 148)
- Road: White (RGB: 255, 255, 255)
- Tree: Dark Green (RGB: 34, 97, 38)
- Water: Blue (RGB: 0, 69, 255)
- Agriculture Land: Light Green (RGB: 75, 181, 73)
- Building: Orange-Red (RGB: 222, 31, 7)
- DeepLabV3+: https://arxiv.org/abs/2210.10732
- OpenEarthMap: https://open-earth-map.org/
- Hugging Face Datasets:
For questions or contributions, feel free to create issues or pull requests.