Ph.D. Thesis · Chulalongkorn University · 2020

FusionNet
GeoLabel

Semantic Segmentation on Remotely Sensed Images Using Deep Convolutional Encoder–Decoder Neural Networks

Teerapong Panboonyuen · Chulalongkorn University

📄 Ph.D. Thesis PDF 💻 GitHub 📑 Defense Slides 📝 Thesis Blog 🏆 ISPRS Leaderboard

90%+

F1-Score on ISPRS Vaihingen

Key Innovations in One Framework

Benchmark Datasets Validated

Project Overview

A research-driven deep learning framework pushing the boundaries of remote sensing scene understanding.

Why does accurate segmentation of remote sensing imagery matter?

Semantic segmentation plays a crucial role in applications spanning agriculture, map updating, urban planning, and disaster response. Yet existing encoder–decoder networks often fail to identify fine low-level features such as rivers and vegetation.

FusionNetGeoLabel proposes an advanced framework specifically designed for aerial and satellite imagery, systematically addressing resolution loss, feature leakage, and limited domain-specific training data.

Remote Sensing Semantic Segmentation Deep Learning Computer Vision Ph.D. Research

Abstract

Five architectural innovations unified in a single framework.

🧠

Global Convolutional Network

Enhances segmentation accuracy by capturing large receptive fields while maintaining classification capability at every spatial position.

🎯

Channel Attention

Focuses network resources on the most discriminative feature channels, improving sensitivity to semantically critical regions.

🔄

Domain-Specific Transfer Learning

Addresses the scarcity of annotated remote sensing data by adapting representations from closely related aerial domains.

🔗

Feature Fusion

Bridges low-level textures with high-level semantics to recover fine details such as rivers, vegetation lines, and road boundaries.

🔭

Depthwise Atrous Convolution

Expands receptive fields at multiple scales efficiently, enabling simultaneous global context and local nuance detection.

📡

HR Backbone Integration

Maintains multi-resolution representations throughout the network, ensuring sharper boundaries and more reliable object detection.

Architecture

HR-GCN-FF-DA: A unified, modular architecture validated on four major benchmarks.

Benchmark Results

FusionNetGeoLabel consistently outperforms prior baselines on ISPRS Vaihingen, Potsdam, Landsat-8, and Massachusetts Roads.

Method	IoU ↑	F1-Score ↑	OA ↑	Dataset
U-Net Baseline	72.4	82.7	85.3	ISPRS Vaihingen
SegNet	74.1	84.3	86.9	ISPRS Vaihingen
GCN + Baseline	78.5	87.2	88.4	ISPRS Vaihingen
GCN + CA + Transfer	82.3	89.5	90.6	ISPRS Vaihingen
HR-GCN-FF-DA (Ours)	86.8	92.1	93.4	ISPRS Vaihingen

Visual Results

Qualitative segmentation outputs on held-out test scenes.

Getting Started

Full semantic segmentation pipeline — train, infer, and evaluate in minutes.

📥 Installation

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

📊 Download Datasets

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

🚀 Training

python train.py --config config.json

🔎 Inference

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

🧪 Evaluation

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

🐳 Docker

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

Selected Publications

Peer-reviewed research derived from this doctoral work.

Transformer-Based Decoder Designs for Semantic Segmentation on Remotely Sensed Images

Panboonyuen, T., et al. · 2021

Remote Sensing, 2021

📄 Read Paper

Feature Fusion-Based Enhanced Global Convolutional Network with Channel Attention for Remote Sensing

Panboonyuen, T., et al. · 2020

Remote Sensing, 2020

📄 Read Paper

Semantic Segmentation on Remotely Sensed Images Using an Enhanced Global Convolutional Network with Channel Attention and Domain Specific Transfer Learning

Panboonyuen, T., et al. · 2019

Remote Sensing, 2019

📄 Read Paper

Road Segmentation on Aerial Imagery Using Deep CNNs and Conditional Random Fields

Panboonyuen, T., et al. · 2017

Remote Sensing, 2017

📄 Read Paper

Academic Journey

Supported by prestigious royal and university scholarships.

Ph.D. in Computer Engineering

Chulalongkorn University, 2018–2020
Thesis: Semantic Segmentation on Remotely Sensed Images Using Deep Convolutional Encoder-Decoder Neural Network

📄 Ph.D. Thesis 📑 Slides 📝 Blog

M.Eng. in Computer Engineering

Chulalongkorn University, 2016–2017
Thesis: High-Resolution Road Extraction Using Deep Convolutional Neural Networks and Conditional Random Fields

📄 M.Eng. Thesis

🏛

The 100th Anniversary Chulalongkorn University Fund for Doctoral Scholarship

🎓

The 90th Anniversary of Chulalongkorn University Scholarship

👑

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

Real-World Impact

Research powering intelligent geospatial systems at scale.

🏙 Urban Planning

Automated aerial scene parsing supports data-driven city development and infrastructure assessment at national scale.

🌿 Environmental Monitoring

Precise vegetation and land-use segmentation enables continuous, large-area ecosystem tracking.

🚨 Disaster Response

Rapid post-event scene understanding accelerates damage assessment and relief coordination in emergency scenarios.

🗺 Geospatial Intelligence

High-accuracy map updating from satellite imagery reduces manual cartographic effort and improves timeliness.

BibTeX Citation

If you use FusionNetGeoLabel in your research, please cite the following.

@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}
}

MIT License · opensource.org/licenses/MIT