> Note: `/app/data` inside the container is the persistent root for accounts, conversations, study areas, the task ledger, and memory — it must be mounted. See the [deployment configuration guide](docs/Deployment.md) for the full variable table.
### Optional: enable "Sign in with ModelScope" (local deployment)
Local deployments also support **one-click sign-in with a ModelScope account** (users skip registration and the ModelScope inference API is configured automatically). Three steps:
1. Sign in at [ModelScope](https://modelscope.cn), go to **Home → Access Control → OAuth Apps → My Apps** and create an application:
- **Redirect URL**: `http://localhost:7860/`;
- **Scopes**: check `profile` and `api-inference` (the rest are optional).
2. Copy the **Client ID** and **Client Secret** into `config/settings.json` of this repository:
```json
"modelscope_oauth": {
"client_id": "your Client ID",
"client_secret": "your Client Secret"
}
```
(Alternatively, inject them via the `MODELSCOPE_OAUTH_CLIENT_ID` / `MODELSCOPE_OAUTH_CLIENT_SECRET` environment variables, which take precedence over the file.)
3. Rebuild and restart — the login page now shows "Sign in with ModelScope": click → authorize on ModelScope → signed in automatically with the inference API configured.
> Note: this repository ships **without any credentials**; do not commit a `config/settings.json` that contains real secrets to a public repository.
## 📸 Feature Gallery
Chat / Work dual mode: clean boundary between Q&A and execution
Approve myself / Full auto: whether key steps wait for your confirmation
When key information is missing (e.g. processing mode), the system asks proactively
Pixel-level temperature query: both 10 m results and 30 m reference are readable pixel by pixel
Accuracy panel: test-area metrics, sample counts, and temperature ranges at a glance
## 🧠 How It Works
A task-control layer sits in front of the algorithms: your goal is decomposed into four responsibilities — **planning, data, training, evaluation** — and orchestrated by agents, while all key computations are delegated to deterministic remote-sensing algorithms. The full pipeline is:
```
Data acquisition → Preprocessing → Terrain thermal response (TTRI) → Random Forest → Thermal constraint correction (TCR) → Export → Accuracy assessment
└ Optional: gap filling (gap-free product)
```
- **Data sources**: Microsoft Planetary Computer (Landsat 8/9 L2) and Copernicus Data Space (Sentinel-2 L2A, GLO-30 DEM), with automatic search, download, caching, and fallback;
- **Temperature source**: Landsat 30 m LST provides the coarse-scale thermal reference; Sentinel-2 provides 10 m spatial detail; Random Forest plus thermal-constraint residual correction yields the 10 m estimate (preserving the 30 m parent-cell mean constraint);
- **Result Q&A**: numerical statistics are computed in code; the LLM only phrases them — **numbers by program, words by model**.
- **Runtime safeguards**: tuning memory pre-check with robust finishing, and failure self-healing (diagnosis → whitelisted action → experience crystallization) are implemented by the task-control layer without changing the deterministic algorithms.
Technical details: [Algorithm Principles](docs/Algorithm.md); for setup and usage, see the [User Guide](docs/User-Guide.md).
## 📖 Documentation
| Document | Content |
|---|---|
| [docs/User-Guide.md](docs/User-Guide.md) | Setup steps, mode descriptions, map & accuracy panels, Wuhan case study |
| [docs/Algorithm.md](docs/Algorithm.md) | Calibration/masking, spectral & terrain features, TTRI, Random Forest, thermal constraint correction, accuracy metrics |
| [docs/Motivation.md](docs/Motivation.md) | Four gaps between raw imagery and usable temperature products |
| [docs/Deployment.md](docs/Deployment.md) | Deployment variable table and migration notes |
## 🛰️ Example Result
Wuhan, 2024-07-22: 10 m LST product (left) vs. 30 m Landsat reference (right)
Compared with the 30 m reference, the 10 m product clearly resolves temperature detail along streets, road networks, and building blocks. The dataset and product for this case are available on [ModelScope](https://modelscope.cn/datasets/Adhara/Wuhan_10m_LST_20240722).
> Important: the downscaled result is not an independently acquired 10 m thermal observation. Test-area accuracy evaluates the model's predictive ability on held-out 30 m temperature samples, and coarse-scale closure evaluates how well the result re-aggregates to the reference mean. Neither substitutes for independent 10 m accuracy validation.
## 📄 Algorithm Inspiration
The algorithm originates from the thermal remote sensing course project of the two developers. Main inspirations:
1. Kolláth, J., Kolláth, M., & Šveda, P. (2022). Combining Landsat 8 and Sentinel-2 data in Google Earth Engine to derive higher resolution land surface temperature maps in urban environment. *Remote Sensing*, 14(16), 4076. doi: 10.3390/rs14164076
2. Yuan, W., Hu, S., Zhan, C., Wang, G., & Luo, Y. (2025). Machine learning land surface temperature downscaling method based on Landsat 9 and Sentinel-2 satellite feature interaction. *Geo-spatial Information Science*, In Press. doi: 10.1080/10095020.2025.2598526
3. Bahi, H., Bounoua, L., Sabri, A., Bannari, A., Malah, A., & Rhinane, H. (2025). A new thermal fusion method to downscale land surface temperature to finer spatial resolution using Sentinel-MSI and Landsat-OLI/TIRS imagery. *Remote Sensing Applications: Society and Environment*, 37, 101519. doi: 10.1016/j.rsase.2025.101519
4. Andriambololonaharisoamalala, R. R., Helmholz, P., Bulatov, D., Ivanova, I., Song, Y., Soon, S., & Jones, E. (2025). Downscaling of urban land surface temperatures using geospatial machine learning with Landsat 8/9 and Sentinel-2 imagery. *Remote Sensing*, 17(14), 2392. doi: 10.3390/rs17142392
5.
## 💬 Contact
GeoThermoAI is developed by two undergraduate students (Class of 2023, Remote Sensing Science and Technology, China University of Geosciences, Wuhan). We warmly welcome collaboration and discussion with researchers in thermal remote sensing.
For bugs or suggestions, please open an Issue or contact us:
- jiyinuo@cug.edu.cn
- 3509851915@qq.com
## ⚖️ License
GeoThermoAI follows a "community edition + commercial license" dual-licensing model: all source code is licensed under **GNU Affero General Public License v3.0 (AGPL-3.0)**. Commercial use requires a separate commercial license — see [NOTICE](NOTICE) for details.
