---
license: mit
pipeline_tag: image-to-image
tags:
- precipitation-nowcasting
- rainmap
- pytorch
- sevir
- fine-tuning
---

# RainMap Nowcasting

This repository hosts model weights for
[`TechieMoon/rainmap-nowcasting`](https://github.com/TechieMoon/rainmap-nowcasting).

## Intended use

The model accepts 6 grayscale rain map frames and predicts the next 6 grayscale
frames. Pixel intensity is interpreted as `0..255 -> 0..1` rain intensity.

## Important warning

Small SEVIR VIL subset fine-tuned benchmark model. Not for operational weather forecasting.

## Model

- Dataset: `SEVIR VIL mini local fine-tune`
- Weights: `rainmap-nowcasting-sevir-finetuned.safetensors`
- Input frames: `6`
- Target frames: `6`
- Image size: `[64, 64]`

## Files

- `rainmap-nowcasting-sevir-finetuned.safetensors`: PyTorch state dict.
- `model_config.json`: architecture, frame count, image size, and encoding.
- `training_metrics.json`: training metrics.


## Benchmark

- Dataset: `data\sevir_mini\local\val`
- Sequences: 4

| Model | MAE | RMSE | CSI@0.3 | HSS@0.3 | ETS@0.3 | FSS@0.3/w15 |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| persistence | 0.0406 | 0.0930 | 0.5944 | 0.7117 | 0.5524 | 0.9675 |
| base | 0.0764 | 0.0983 | 0.6432 | 0.7520 | 0.6026 | 0.9605 |
| fine_tuned | 0.0694 | 0.0888 | 0.6699 | 0.7747 | 0.6323 | 0.9727 |

Small-subset benchmark results are not operational forecasting claims.


## Example

```bash
python -m rainmap_nowcasting.predict --input-dir samples/input --output-dir outputs
```