Task-Driven Dictionary Learning with Deep Metric Learning for Remote Sensing Scene Classification

This repository provides an implementation of a Task-Driven Dictionary Learning (TDDL) framework integrated with deep metric learning in a triplet architecture for high-resolution remote sensing scene classification.

The method combines:

• Superpixel-based region decomposition (SLIC),
• Region-level feature extraction,
• Graph-based representation learning (GCN + attention),
• Dictionary learning optimized for classification objectives,
• Metric learning through triplet loss to improve class separability.

---

Overview

High-resolution remote sensing images often contain complex spatial patterns and intra-class variability.
This project addresses these challenges by coupling dictionary learning with graph neural modeling and metric learning.

At a high level, the pipeline is:

1. Segment each image into regions using SLIC superpixels.
2. Extract local region descriptors.
3. Construct graph-structured representations.
4. Pretrain a graph model (GCN + attention).
5. Train the full TDDL-based triplet framework for discriminative embedding learning.
6. Perform scene classification from learned representations.

---

Repository Structure

Update this section if your folder/file names differ.

• train_tddl_gcn.py — demo training script for the 2-stage training process.
• models/ — model definitions (GCN, attention modules, triplet architecture).
• data/ — dataset loaders and preprocessing utilities.
• utils/ — helper functions for training/evaluation.
• README.md — project documentation.

---

Datasets

The implementation is designed for the following remote sensing scene datasets:

• UC Merced Land Use Dataset (UCMD)
  http://weegee.vision.ucmerced.edu/datasets/landuse.html
• Aerial Image Dataset (AID)
  https://captain-whu.github.io/AID/
• PatternNet
  https://drive.google.com/file/d/127lxXYqzO6Bd0yZhvEbgIfz95HaEnr9K/view

Please download datasets from their official sources and organize them according to your local data loading configuration.

---

Preprocessing

As a preprocessing step, each image is partitioned into regions in an unsupervised way using the SLIC algorithm.

For each region:

• local descriptors are extracted,
• descriptors are used to build graph-structured inputs for the network.

If external MATLAB feature extraction scripts are required, ensure they are available and correctly referenced in your preprocessing pipeline.

---

Training

A demo training entry point is provided in:

• train_tddl_gcn.py

The current training procedure is configured as a two-stage process:

1. Pretraining stage: train a base GCN + attention model.
2. Task-driven stage: optimize the full dictionary learning + metric learning objective using triplet supervision.

To run training (example):

python train_tddl_gcn.py

Adjust command-line arguments or configuration values according to your dataset paths, hyperparameters, and environment.

---

Requirements

Create and activate a Python environment, then install required packages.

pip install -r requirements.txt

If requirements.txt is not yet complete, include at least:

• torch
• torchvision
• numpy
• scikit-learn
• opencv-python
• scikit-image
• networkx (if used by graph construction code)
• additional dataset/preprocessing dependencies as needed

---

Reproducibility Notes

For reproducible experiments, consider fixing:

• random seeds,
• train/val/test split definitions,
• superpixel parameters,
• dictionary size and sparse coding settings,
• triplet mining strategy and margin values.

---

Citation

If you use this repository in your research, please cite:

@article{Regan2023TripletGCN,
  title   = {A triplet graph convolutional network with attention and similarity-driven dictionary learning for remote sensing image retrieval},
  author  = {Regan, Jacob and Khodayar, Mahdi},
  journal = {Expert Systems with Applications},
  volume  = {232},
  pages   = {120579},
  year    = {2023},
  issn    = {0957-4174},
  doi     = {10.1016/j.eswa.2023.120579},
  url     = {https://www.sciencedirect.com/science/article/pii/S0957417423010813}
}

---

Acknowledgements

This project builds on ideas from:

• task-driven dictionary learning,
• graph neural networks,
• deep metric learning for visual representation learning,
• publicly available remote sensing datasets and preprocessing tools.

---

License

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