GeoNeXt: Efficient Landslide Mapping using ConvNeXt V2 and PSA-ASPP

GeoNeXt: Efficient landslide mapping using a pre-trained ConvNeXt V2 encoder with a PSA-ASPP decoder

Rodrigo Uribe-Ventura, Willem Viveen, Ferdinand Pineda-Ancco, César Beltrán-Castallon

Overview

GeoNeXt is a novel deep learning architecture for landslide detection and mapping that combines a pre-trained ConvNeXt V2 encoder with a decoder integrating Pyramid Squeeze Attention (PSA) and Atrous Spatial Pyramid Pooling (ASPP). The model achieves state-of-the-art performance with 10× fewer parameters than transformer-based approaches while maintaining superior accuracy across diverse geographical regions.

Key Features

• High Performance: F1 scores of 94.25%, 86.43%, and 92.27% on Bijie, Landslide4Sense, and GVLM datasets
• Computational Efficiency: 32.2M parameters vs 312.5M for SAM-based methods
• Zero-shot Transferability: Strong cross-domain performance without retraining
• Multi-scale Detection: Handles landslides from small flows to large-scale events

Architecture

GeoNeXt follows a U-Net-based encoder-decoder structure:

• Encoder: Pre-trained ConvNeXt V2-Tiny with domain adaptation
• Decoder: Novel PSA-ASPP design for multi-scale feature aggregation
• Loss Function: Composite loss combining BCE, uncertainty-weighted Dice, and focal-Tversky

Installation

Setup

# Clone the repository
git clone https://github.com/ruribev/GeoNeXt.git
cd GeoNeXt

# Install dependencies
pip install -r requirements.txt

Download Required Resources

Download the following required folders from Google Drive:

• Pre-trained Weights: Download pretrained folder
• Trained Models: Download experiments folder
• Datasets: Download dataset folder

Extract these folders in the root directory of the project to maintain the proper structure.

Quick Start

Training

from utils.experiments import TrainingExperiment, run_experiment

# Configure training experiment
config = {
    'dataset': 'BJL',  # or 'L4S', 'GVLM', 'CAS'
    'model_variant': 'tiny',
    'gpu_id': 0,
    'batch_size': 8,
    'epochs': 60,
    'lr': 3.5e-3,
    'convnextv2_pretrained_weights': 'pretrained/convnextv2_tiny_22k_224_ema.pt',
    'name_suffix': 'your_experiment_name'
}

# Run training
experiment = TrainingExperiment(config)
run_experiment(experiment)

Benchmarking

from utils.benchmarks import Benchmark

# Initialize benchmark
benchmark = Benchmark(
    dataset='BJL',
    gpu_id=0,  
    model_variant='tiny',
    model_weight_path="experiments/GeoNeXt_BJL/best_model.pth.tar",
    batch_size=2
)

# Run evaluation
results = benchmark.run_benchmark()
benchmark.show_metrics()
benchmark.show_examples(num_examples=5, min_mask_percentage=0.1)

Jupyter Notebooks

We provide ready-to-use Jupyter notebooks:

• train.ipynb: Complete training pipeline example
• benchmark.ipynb: Evaluation of main datasets

Data Preparation

1. Download the complete dataset folder from this link
2. Extract and place in the project root directory
3. The structure should be:

dataset/
├── BJL/
├── L4S/
├── GVLM/
└── CAS/

Project Structure

GeoNeXt/
├── models/
│   ├── __init__.py
│   ├── GeoNeXt.py          # Main model architecture
│   └── ConvNeXtV2.py       # ConvNeXt V2 backbone
├── utils/
│   ├── __init__.py
│   ├── train.py            # Training utilities
│   ├── evaluator.py        # Evaluation metrics
│   ├── benchmarks.py       # Benchmarking tools
│   ├── experiments.py      # Experiment management
│   ├── augmentations.py    # Data augmentation
│   ├── losses.py           # Loss functions
│   ├── ema.py              # Exponential moving average
│   └── dataset.py          # Dataset utilities
├── dataset/                # Dataset directory (download from link above)
├── pretrained/             # Pre-trained weights (download from link above)
├── experiments/            # Training outputs (download from link above)
├── train.ipynb            # Training notebook
├── benchmark.ipynb        # Evaluation notebook
├── requirements.txt       # Dependencies
├── LICENSE                # MIT License
└── README.md             # This file

Pre-trained Models & Weights

Available Model Weights

Pre-trained models and weights are available in the downloads section:

• ConvNeXt V2 Tiny Backbone: ImageNet-22K pre-trained weights (32.2M parameters)

  • File: convnextv2_tiny_22k_224_ema.pt
  • Location: pretrained folder

• GeoNeXt Complete Models: Fine-tuned models for each dataset

  • BJL Dataset: GeoNeXt_BJL/best_model.pth.tar (F1: 94.25%)
  • L4S Dataset: GeoNeXt_L4S/best_model.pth.tar (F1: 86.43%)
  • GVLM Dataset: GeoNeXt_GVLM/best_model.pth.tar (F1: 92.27%)
  • CAS Domain-adapted: GeoNeXt_CAS/encoder_weights.pth
  • Location: experiments folder

Download the pretrained folder and experiments folder to access all pre-trained models and weights.

License

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

Citation

If you use GeoNeXt in your research, please cite our work:

@article{uribe2025geonext,
  title={GeoNeXt: Efficient landslide mapping using a pre-trained ConvNeXt V2 encoder with a PSA-ASPP decoder},
  author={Uribe-Ventura, Rodrigo and Viveen, Willem and Pineda-Ancco, Ferdinand and Beltr{\'a}n-Casta{\~n}on, C{\'e}sar},
  journal={Artificial Intelligence in Geosciences},
  year={2025},
  publisher={Elsevier},
  doi={10.1016/j.aiig.2025.100172}
}

Paper currently under review. BibTeX entry will be updated upon publication.

Contact

• Corresponding Author: Rodrigo Uribe-Ventura (a20234215@pucp.edu.pe)
• Institution: Pontificia Universidad Católica del Perú
• GitHub: @ruribev

---

For detailed methodology and experimental results, please refer to our paper: "GeoNeXt: Efficient landslide mapping using a pre-trained ConvNeXt V2 encoder with a PSA-ASPP decoder"