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.
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:
git clone https://huggingface.co/Jamtang/classification
cd classification
2. Create a Python virtual environment and install dependencies:
conda create -n class python=3.8 -y
conda activate class
pip install -r requirements.txt
🚀 Usage
Inference / Application
To run the provided application:
python app.py
Training
To train the model on a specific dataset (e.g., AID):
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
Citation
If you find this work helpful or inspiring, please consider citing it:
@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}
}