---
title: TerraScope-Segformer
canonical_url: "https://www.modelscope.cn/models/LazySheep006/TerraScope-Segformer"
md_url: "https://www.modelscope.cn/models/LazySheep006/TerraScope-Segformer.md"
repository: LazySheep006/TerraScope-Segformer
last_updated: 2026-09-23
license: other
pipeline_tag: image-segmentation
tasks:
  - image-segmentation
base_model:
  - nvidia/mit-b5
base_model_relation: finetune
downloads: 3
stars: 0
tags:
  - remote-sensing
  - semantic-segmentation
  - segformer
  - mmsegmentation
---

# TerraScope-Segformer

> TerraScope-Segformer - LazySheep006 在 ModelScope 开源的模型。基于Segformer的遥感语义分割垂直微调

LazySheep006/TerraScope-Segformer 是 ModelScope 魔搭社区上的image-segmentation模型，采用 other 许可，基于 nvidia/mit-b5 构建。

- **Repository**: LazySheep006/TerraScope-Segformer
- **License**: other
- **Tasks**: image-segmentation
- **Base model**: nvidia/mit-b5
- **Tags**: remote-sensing, semantic-segmentation, segformer, mmsegmentation
- **Downloads**: 3
- **Stars**: 0
- **Last updated**: 2026-09-23

Source: https://www.modelscope.cn/models/LazySheep006/TerraScope-Segformer

---

# TerraScope Segmentation

> A SegFormer-B5 model for semantic segmentation of remote-sensing imagery.

TerraScope Segmentation is a seven-class semantic-segmentation model based on SegFormer-B5 and fine-tuned on LoveDA. It produces dense, pixel-level predictions for common land-cover categories in urban and rural remote-sensing scenes. The model is provided as a standalone MMSegmentation checkpoint with its inference configuration.

LoveDA adaptation by [Ningkai Wu](https://github.com/kiny007).

Model release: [LazySheep006/TerraScope-Segformer](https://modelscope.cn/models/LazySheep006/TerraScope-Segformer)

## Model Details

| Property | Value |
| --- | --- |
| Architecture | SegFormer-B5 |
| Backbone | MixVisionTransformer-B5 |
| Task | Semantic segmentation |
| Dataset | LoveDA |
| Number of classes | 7 |
| Inference framework | MMSegmentation 1.2.2 |
| Checkpoint format | PyTorch `.pth` |
| Best checkpoint | `best_mIoU_iter_10000.pth` |
| Validation mIoU | 53.85 |

## Classes

Class indices and palette entries follow the order below:

| Index | Class | Palette |
| ---: | --- | --- |
| 0 | Background | `(255, 255, 255)` |
| 1 | Building | `(255, 0, 0)` |
| 2 | Road | `(255, 255, 0)` |
| 3 | Water | `(0, 0, 255)` |
| 4 | Barren | `(159, 129, 183)` |
| 5 | Forest | `(0, 255, 0)` |
| 6 | Agricultural | `(255, 195, 128)` |

## Inference

The model directory contains both files required for standalone inference:

```text
config.py
checkpoint.pth
```

Install a PyTorch and MMCV build compatible with the target CUDA runtime, followed by MMSegmentation:

```bash
pip install mmengine==0.10.7 mmsegmentation==1.2.2
```

```python
from mmseg.apis import inference_model, init_model

model = init_model(
    "config.py",
    "checkpoint.pth",
    device="cuda:0",
)

result = inference_model(model, "/path/to/image.png")
mask = result.pred_sem_seg.data.squeeze(0).cpu().numpy()
print(mask.shape)
```

Each value in `mask` is a class index from the table above. The accompanying configuration applies RGB normalization and preserves the declared class order and palette.

## Intended Use

TerraScope Segmentation is intended for non-commercial research and evaluation in remote-sensing image analysis. Performance can vary with sensor characteristics, geographic region, ground sampling distance, atmospheric conditions, image compression, and domain shift. Predictions should be independently verified before consequential use.

## License and Acknowledgements

The original SegFormer project is distributed under the [NVIDIA Source Code License for SegFormer](https://github.com/NVlabs/SegFormer/blob/master/LICENSE), which restricts the work and its derivatives to non-commercial research or evaluation and requires the complete license to accompany redistribution. A copy is included with the model as [`LICENSE`](LICENSE).

The OpenMMLab MMSegmentation implementation is separately released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). LoveDA and other third-party assets remain subject to their respective licenses and terms. The NVIDIA license remains the controlling restriction wherever the SegFormer work or its derivatives apply.

Upstream repositories: [NVlabs/SegFormer](https://github.com/NVlabs/SegFormer) · [open-mmlab/mmsegmentation](https://github.com/open-mmlab/mmsegmentation)