CSPS_RetinaNet: An Enhanced RetinaNet for Rodent Burrow Detection in Grassland Ecosystems

Overview

CSPS_RetinaNet is an improved object detection model based on RetinaNet, specifically designed for detecting rodent burrows in grassland ecosystems using high-resolution satellite imagery. This repository contains the implementation of our novel architecture that integrates CSPDarknet53 backbone with SimAM attention mechanism and GIoU loss for enhanced detection performance.

Contributors

Haitao Sun^a, Songbing Zou^a*, Zhenqing Ji^a, Yajie Bai^a, Wenyong Zhang^a, Tenghao Gou^a, Pengxiang Xie^a

^a Key Laboratory of Western China's Environmental Systems (Ministry of Education), College of Earth and Environmental Sciences, Lanzhou University, Lanzhou, 730000, China.

*Corresponding author

Key Features

• CSPDarknet53 Backbone: Efficient feature extraction with cross-stage partial connections
• SimAM Attention Mechanism: Parameter-free spatial attention module for enhanced feature representation
• GIoU Loss: Improved bounding box regression for better localization accuracy
• Feature Pyramid Network (FPN): Multi-scale feature fusion for detecting burrows of various sizes

Model Architecture

The CSPS_RetinaNet architecture consists of:

1. Backbone Network: CSPDarknet53 for robust feature extraction
2. Attention Module: SimAM applied at C3 feature level
3. Neck: Feature Pyramid Network (FPN) generating P3-P7 feature levels
4. Detection Heads: Separate classification and regression branches

Installation

Requirements

Python >= 3.7
PyTorch >= 1.7.1
CUDA >= 10.2 (for GPU support)

Dependencies

pip install -r requirements.txt

Required packages:

• torch
• torchvision
• numpy
• opencv-python
• pillow
• matplotlib
• scipy
• tqdm
• pycocotools

Dataset

Data Structure

Organize your dataset in VOC format:

VOCdevkit/
└── VOC2007/
    ├── Annotations/     # XML annotation files
    ├── JPEGImages/      # Image files
    └── ImageSets/
        └── Main/
            ├── train.txt
            ├── val.txt
            └── test.txt

Data Preparation

1. Prepare your images and annotations in VOC format
2. Run the annotation script to generate training files:

python voc_annotation.py

This will generate 2007_train.txt and 2007_val.txt for training.

Configuration

Model Configuration

Edit model_data/cls_classes.txt to define your class names (one class per line).

Training Configuration

Key parameters in train.py:

• input_shape: [600, 600] - Input image size
• phi: 5 - Use CSPDarknet53 backbone
• Init_lr: 1e-4 - Initial learning rate (Adam optimizer)
• Freeze_Epoch: 60 - Epochs for frozen training
• UnFreeze_Epoch: 120 - Total training epochs
• batch_size: Adjust based on your GPU memory

Usage

Training

1. Prepare your dataset following the VOC format structure

2. Configure training parameters in train.py:

   • Set classes_path to your classes file
   • Set model_path to pretrained weights (optional)
   • Adjust batch_size based on GPU memory
   • Configure training epochs

3. Start training:

python train.py

The model will automatically:

• Download pretrained backbone weights
• Perform frozen training (backbone frozen)
• Perform unfrozen training (full model training)
• Save checkpoints in logs/ directory
• Generate loss curves and evaluation metrics

Prediction

For single image prediction:

from retinanet import Retinanet

# Initialize model
model = Retinanet()

# Predict on image
from PIL import Image
image = Image.open('your_image.jpg')
result = model.detect_image(image)
result.show()

For batch prediction:

# Set mode to "dir_predict" in predict.py
python predict.py

Evaluation

Calculate mAP on test set:

python get_map.py

This will generate:

• Detection results for each image
• Precision-Recall curves
• mAP scores
• F1 scores, Recall, and Precision metrics

Model Performance

The model achieves state-of-the-art performance on rodent burrow detection:

• Optimized for high-resolution (3000×3000) Sentinel-2 imagery
• Robust detection of small objects (burrows)
• Improved localization accuracy with GIoU loss

Advanced Features

Mixed Precision Training (FP16)

Enable FP16 for faster training with reduced memory:

fp16 = True  # in train.py

Multi-GPU Training

For distributed training:

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 train.py

Set distributed = True in train.py.

Model Summary

View model architecture and parameters:

python summary.py

Data Availability

• Rodent Burrow Dataset: Available at https://zenodo.org/records/17491180
• Trained Model Weights: Available at https://zenodo.org/records/17491180
• Implementation Code: Available at https://github.com/yeshushahua/CSPS_RetinaNet

Citation

If you use this code or dataset in your research, please cite:

@article{sun2025csps,
  title={CSPS_RetinaNet: An Enhanced RetinaNet for Rodent Burrow Detection in Grassland Ecosystems},
  author={Sun, Haitao and Zou, Songbing and Ji, Zhenqing and Bai, Yajie and Zhang, Wenyong and Gou, Tenghao and Xie, Pengxiang},
  journal={},
  year={2025},
  publisher={}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

For questions and support, please contact:

• Songbing Zou (Corresponding Author): [email]
• Open an issue on GitHub: https://github.com/yeshushahua/CSPS_RetinaNet/issues

---

Last Updated: November 2025