🌍 FusionNetGeoLabel

📚 Project Overview

FusionNetGeoLabel is a research-driven deep learning framework for semantic segmentation of remote sensing imagery. Built upon my doctoral research, it introduces a novel architecture that integrates:

• High-Resolution Backbone (HR): preserves spatial precision by maintaining multi-resolution representations throughout the network.
• Feature Fusion (FF): bridges low-level and high-level features to recover fine details such as rivers, vegetation, and road boundaries.
• Depthwise Atrous Convolution (DA): enhances receptive fields at multiple scales while reducing parameter overhead.

This unified design, named HR-GCN-FF-DA, delivers state-of-the-art performance on benchmarks such as ISPRS Vaihingen, Potsdam, Landsat-8, and Massachusetts Roads, consistently surpassing existing baselines in IoU, F1-score, and overall accuracy.

Beyond academic evaluation, FusionNetGeoLabel is designed with practical deployment in mind — featuring modular utilities, dataset download scripts, pretrained models, and Docker support for seamless reproducibility. The goal is to provide both a research-grade contribution to the remote sensing community and a production-ready tool for real-world geospatial applications such as urban planning, agriculture monitoring, and map updating.

🎓 Academic Journey & Scholarships

I received my Ph.D. in Computer Engineering from Chulalongkorn University (2018–2020), supported by two prestigious scholarships:

• The 100th Anniversary Chulalongkorn University Fund for Doctoral Scholarship
• The 90th Anniversary of Chulalongkorn University Scholarship

Prior to this, I received my Master of Engineering in Computer Engineering from Chulalongkorn University (2016–2017), supported by:

• H.M. the King Bhumibol Adulyadej’s 72nd Birthday Anniversary Scholarship

📖 Thesis Works

📝 Selected Publications

My research contributes to the advancement of intelligent systems in geospatial analysis — supporting smart cities, environmental monitoring, disaster response, and geospatial intelligence with more robust and accurate semantic segmentation models.

📄 Abstract

Semantic segmentation plays a crucial role in remote sensing, impacting fields such as agriculture, map updating, and navigation.

While Deep Convolutional Encoder-Decoder networks are widely used, they often struggle to accurately identify fine low-level features such as rivers and vegetation due to architectural limits and scarcity of domain-specific training data.

This dissertation proposes an advanced semantic segmentation framework designed specifically for remote sensing imagery, featuring five key innovations:

• Global Convolutional Network (GCN): Enhances segmentation accuracy for remote sensing images.
• Channel Attention Mechanism: Focuses on the most critical features for better performance.
• Domain-Specific Transfer Learning: Addresses limited training data challenges.
• Feature Fusion (FF): Integrates low-level features effectively.
• Depthwise Atrous Convolution (DA): Refines feature extraction for improved segmentation.

Experiments on Landsat-8 datasets and the ISPRS Vaihingen benchmark demonstrate significant performance improvements over baseline models.

📁 Key Resources & Publications

Explore the core assets underpinning my research and contributions to the field of semantic segmentation on remote sensing imagery:

These resources highlight the rigor, reproducibility, and impact of my work within the computer vision and remote sensing communities.

🔧 Getting Started

The FusionNetGeoLabel framework provides a full semantic segmentation pipeline built on HR-Backbone + Feature Fusion + Depthwise Atrous Convolution. Below are quick steps to run training, inference, and evaluation.

📥 Installation

git clone https://github.com/kaopanboonyuen/FusionNetGeoLabel.git
cd FusionNetGeoLabel
pip install -r requirements.txt

📊 Training

Prepare your dataset (ISPRS Vaihingen, Potsdam, Massachusetts Roads, or Landsat-8) using our dataset download scripts:

bash scripts/download_isprs.sh
bash scripts/download_potsdam.sh
bash scripts/download_massachusetts.sh

Modify config.json (e.g., dataset paths, hyperparameters), then start training:

python train.py --config config.json

🔎 Inference

Run inference on single images or folders using pretrained checkpoints:

python inference.py --model checkpoints/hrgcn_ff_da.pth --image sample.png

🧪 Evaluation

Test your model against benchmark datasets with built-in metrics (IoU, F1-score, Accuracy):

python test.py --config config.json --model checkpoints/hrgcn_ff_da.pth

🐳 Docker Support

Build and run inside a container for reproducibility:

docker build -t fusionnetgeolabel .
docker run --gpus all -it fusionnetgeolabel

📸 Visual Results

Some highlights of our model's performance:

🚀 What I Do & My Impact

I design and develop advanced deep learning architectures for semantic segmentation of aerial and satellite imagery, enabling machines to interpret complex geospatial scenes — from roads and vegetation to urban structures — with unprecedented precision.

My latest framework, FusionNetGeoLabel (HR-GCN-FF-DA), pushes beyond the state of the art by introducing three key innovations:

• High-Resolution Backbone: Maintains fine-grained spatial representations across all layers, ensuring sharper boundaries and more reliable object detection.
• Feature Fusion: Bridges low-level textures with high-level semantics, capturing subtle details (e.g., rivers, vegetation lines) that standard networks often miss.
• Depthwise Atrous Convolution: Expands receptive fields efficiently at multiple scales, enabling the network to detect both global context and local nuances.

Validated on leading benchmarks such as ISPRS Vaihingen and Landsat-8, FusionNetGeoLabel consistently achieves 90%+ F1-scores — surpassing previous baselines and setting new standards in remote sensing segmentation.

Beyond research impact, my work powers practical applications in urban planning, environmental monitoring, disaster management, and navigation systems, directly contributing to smarter, data-driven decision-making at scale.

⚖️ License

This project is licensed under the MIT License.

📝 Citation

If you use this framework, please cite the following thesis:

@phdthesis{panboonyuen2019semantic,
  title     = {Semantic segmentation on remotely sensed images using deep convolutional encoder-decoder neural network},
  author    = {Teerapong Panboonyuen},
  year      = {2019},
  school    = {Chulalongkorn University},
  type      = {Ph.D. thesis},
  doi       = {10.58837/CHULA.THE.2019.158},
  address   = {Faculty of Engineering},
  note      = {Doctor of Philosophy}
}