---
license: apache-2.0
base_model: Qwen/Qwen3-VL-4B-Instruct
library_name: peft
pipeline_tag: image-text-to-text
language:
- en
tags:
- remote-sensing
- temporal-understanding
- multimodal
- lora
- earth-observation
- qwen3-vl
---

# GeoChrono

**GeoChrono** is a multimodal large language model (MLLM) tailored for long-term remote sensing interpretation. Given an image sequence spanning multiple years and a text prompt, GeoChrono generates a response through spatio-temporal reasoning — tracing, memorizing, and reasoning about the long-term evolution of geographic entities.

GeoChrono is introduced in the paper [*GeoChrono: Benchmarking and Rethinking Long-Term Temporal Understanding in Remote Sensing*](https://arxiv.org/abs/2607.15768) (**accepted at ACM MM 2026** 🔥), together with the **ChronoBench** benchmark and the **ChronoInstruct** instruction-tuning dataset.

- 📄 **Paper**: https://arxiv.org/abs/2607.15768
- 🌐 **Project Page**: https://davidup1.github.io/papers/geochrono/
- 📦 **Code**: https://github.com/IntelliSensing/GeoChrono
- 📊 **Data (ChronoBench & ChronoInstruct)**: https://huggingface.co/datasets/Davidup1/GeoChrono-Data

## Model Architecture

GeoChrono comprises four core components: **i)** a Visual Encoder that extracts per-frame visual features, **ii)** a Vision-Language Projector that maps visual representations into the language embedding space, **iii)** a **Temporal Trajectory Encoder (TempEnc)** that models per-location temporal evolution to strengthen the perception of land-cover change history, and **iv)** a Large Language Model (LLM) that performs multimodal reasoning and generates the textual response.

**TempEnc** leverages the physical prior that each geographic parcel remains spatially fixed while its semantics evolve: it decouples the spatio-temporal feature volume into per-location temporal trajectories via three stages — Spatial Context Aggregation (a lightweight 3×3 depthwise convolution), Hybrid Temporal Attention (dual-stream bidirectional + causal self-attention along the temporal axis), and Semantic Focusing (text-guided cross-attention that suppresses task-irrelevant temporal information).

The companion **Coarse-to-Fine Token Compressor (C2FComp)** leverages prompt text embeddings to assess the task relevance of each spatial region, selectively preserving full-resolution fine tokens for salient areas while condensing the static background into compact coarse representations — reducing visual tokens by over 56% while retaining 94.6% of the full model's performance. C2FComp is available in the code repository.

## Training

GeoChrono is built upon **Qwen3-VL-4B-Instruct**. During training, the Vision Encoder and Vision-Language Projector are frozen; TempEnc is randomly initialized and fully fine-tuned, while the LLM backbone is tuned with LoRA (r=32, α=64, on `q_proj`/`k_proj`/`v_proj`/`o_proj`). TempEnc adopts a single-layer attention architecture. The model is trained for one epoch on ChronoInstruct (104K samples) using 4 NVIDIA H100 80GB GPUs. All images are fed at a fixed 1024×1024 resolution.

## Repository Contents

| File | Description |
|:---|:---|
| `adapter_model.safetensors` / `adapter_config.json` | LoRA adapter for the Qwen3-VL-4B-Instruct LLM backbone |
| `ottc_weights.pt` / `ottc_config.json` | TempEnc weights and configuration (`ottc` is the internal codename of TempEnc) |
| `tokenizer.json`, `preprocessor_config.json`, ... | Tokenizer and processor files (identical to the base model) |

## Usage

GeoChrono requires the custom TempEnc injection code from the GitHub repository — it cannot be loaded with vanilla `transformers` + `peft` alone:

```bash
git clone https://github.com/IntelliSensing/GeoChrono.git
```

Follow the inference / evaluation instructions in the repository README: the base model is `Qwen/Qwen3-VL-4B-Instruct`, on top of which the LoRA adapter is applied and the TempEnc module (`ottc_weights.pt`) is injected into the model forward pass.

## Results on ChronoBench

GeoChrono achieves state-of-the-art performance on ChronoBench, surpassing the leading commercial MLLMs by over 20% and reaching human-level accuracy on several sub-tasks:

| Method | LCP | TR | LTM | STR | OA |
|:---|---:|---:|---:|---:|---:|
| Human | 97.04 | 89.78 | 91.73 | 95.56 | 92.28 |
| Gemini-3-Flash | 65.51 | 61.38 | 47.52 | 59.89 | 57.48 |
| GPT-5.4 | 43.53 | 67.57 | 39.21 | 50.42 | 56.29 |
| Qwen3-VL-32B | 43.76 | 57.88 | 24.06 | 47.23 | 46.73 |
| DVLChat-4B | 47.74 | 34.21 | 22.91 | 40.59 | 44.07 |
| **GeoChrono (ours)** | **88.65** | **83.03** | **68.10** | **72.92** | **78.34** |

(LCP: Land Cover Perception, TR: Temporal Recognition, LTM: Long-Term Memory, STR: Spatio-Temporal Reasoning, OA: Overall Accuracy. All values in %.)

GeoChrono also generalizes to unseen distributions: under zero-shot settings it reaches 72.9 / 42.1 / 35.7 on the three DVL-Bench sub-tasks and 59.2 on CDVQA, outperforming all remote sensing domain baselines.

## License

This model is released under the **Apache License 2.0**. The base model Qwen3-VL-4B-Instruct is subject to its own license.

## Citation

If you find GeoChrono useful, please cite our paper:

```bibtex
@article{li2026geochrono,
  title   = {GeoChrono: Benchmarking and Rethinking Long-Term Temporal Understanding in Remote Sensing},
  author  = {Li, Yujie and Pan, Jiancheng and Wei, Zhiwei and Wang, Jiuniu and Peng, Mugen and Xu, Wenjia},
  journal = {arXiv preprint arXiv:2607.15768},
  year    = {2026}
}
```

## Contact

For questions or feedback, please open a discussion or contact **liyujie2003@bupt.edu.cn**.