File size: 3,019 Bytes
32a1a30
 
e5fbb03
 
 
 
 
 
32a1a30
e5fbb03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
964d3cb
f253d86
 
964d3cb
e5fbb03
 
 
 
 
 
 
 
 
 
 
 
964d3cb
d78f40e
2a18121
690cea3
e5fbb03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
---
license: apache-2.0
pipeline_tag: image-classification
tags:
- remote-sensing
- mamba
- mixture-of-experts
- cnn
---

# AFM-Net: Hierarchical Fusion of Local and Global Visual Features with Mixture-of-Experts for Remote Sensing Image Scene Classification

This repository contains the pre-trained model weights for **AFM-Net**, presented in the paper [Hierarchical Fusion of Local and Global Visual Features with Mixture-of-Experts for Remote Sensing Image Scene Classification](https://huggingface.co/papers/2510.27155).

AFM-Net is a parallel heterogeneous framework that synergizes CNN-based local texture extraction with Mamba-based global sequence modeling. It utilizes a hierarchical fusion module to integrate features across different semantic levels and an adaptive Mixture-of-Experts (MoE) classifier head for fine-grained scene recognition.

## 🔍 Introduction 

Remote sensing scene classification of high-resolution images remains a challenging task due to complex spatial structures and high intra-class variance. We propose a parallel heterogeneous framework that:
- **Synergizes Local and Global Visual Encoder**: Coupling CNN-based local texture extraction with Mamba-based global sequence modeling to ensure robust co-representation.
- **Hierarchical Fusion Strategy**: Employs a hierarchical fusion module to densely integrate heterogeneous features across different semantic levels.
- **Adaptive MoE Classifier**: Uses a Mixture-of-Experts (MoE) head to dynamically select optimal features, balancing performance with efficiency.

Extensive experiments show our model achieves state-of-the-art performance on AID, NWPU-RESISC45, and UC Merced datasets.

## ⚙️ Installation 

### 1. Clone this repository:
```bash
git clone https://huggingface.co/Jamtang/classification
cd classification
```

### 2. Create a Python virtual environment and install dependencies:
```bash
conda create -n class python=3.8 -y
conda activate class
pip install -r requirements.txt
```

## 🚀 Usage 

### Inference / Application
To run the provided application:
```bash
python app.py
```

### Training 
To train the model on a specific dataset (e.g., AID):
 ```bash
python train.py --dataset AID --batch_size 32 --epochs 500
 ```

## 🔥 Performance 
| Dataset | F1 Score |
| :--- | :---: |
| UC Merced | 96.81 |
| AID | 93.71 |
| NWPU-RESISC45 | 95.52 |

## Links
- **Paper**: [https://huggingface.co/papers/2510.27155](https://huggingface.co/papers/2510.27155)
- **Code**: [https://github.com/tangyuanhao-qhu/AFM-Net](https://github.com/tangyuanhao-qhu/AFM-Net)

## Citation
If you find this work helpful or inspiring, please consider citing it:
```bibtex
@article{tang2025hierarchical,
  title={Hierarchical Fusion of Local and Global Visual Features with Mixture-of-Experts for Remote Sensing Image Scene Classification},
  author={Tang, Yuanhao and Zou, Xuechao and Hu, Zhengpei and Xing, Junliang and Zhang, Chengkun and Huang, Jianqiang},
  journal={arXiv preprint arXiv:2510.27155},
  year={2025}
}
```