GLADE-Net

learning framework for perceptual dehazing of high-resolution remote sensing images: Phase 1 (Global Attention Stage): Uses a UFormer-based encoder–decoder with grid patching to capture global atmospheric and ecological context. Phase 2 (Laplacian-Guided GAN Stage): Refines outputs via a dual-branch GAN that operates in both RGB and Laplacian domains, enhancing edges, textures, and perceptual realism. GLADE-Net achieves state-of-the-art PSNR and LPIPS on SateHaze1k and RICE datasets, balancing structural fidelity and visual quality for environmental and ecological monitoring applications.

---

🧠 Overview

Phase 1 — Global Attention Dehazing

This stage focuses on coarse haze removal and global feature restoration using a UFormer-based encoder–decoder optimized for remote sensing imagery.

Key Features:

• 🌍 Global Attention Encoder–Decoder (UFormer): Captures large-scale atmospheric and ecological dependencies.
• ⚡ PyTorch Lightning Framework: Streamlined training loop with mixed-precision (FP16) support.
• 🧩 Grid-Based Dataset Loader: Performs on-the-fly tiling, augmentation, and efficient handling of high-resolution inputs.
• 📈 Gradual Frequency Loss: Charbonnier-based loss with PSNR/SSIM evaluation metrics for stable optimization.
• 🛰️ Inference Tools: Enable large-scale aerial image restoration through automated grid merging and tiling scripts.

---

Phase 2 — Laplacian-Guided Perceptual Refinement

This stage enhances visual quality and perceptual realism of the dehazed outputs generated by Phase 1.
It integrates adversarial learning and frequency-domain guidance for fine detail enhancement.

Key Features:

• 🎨 Dual-Branch GAN: Operates jointly in RGB and Laplacian domains to restore both structural and textural details.
• 🔁 Laplacian Pyramid Guidance: Multi-scale refinement of edges and textures using Laplacian-based supervision.
• 👁️ Perceptual and Adversarial Losses: Improve realism and visual fidelity while maintaining quantitative accuracy.
• 🔬 High-Resolution Enhancement: Designed for ecological monitoring and environmental analysis applications.

---

⚙️ Installation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Tested with Python 3.8+ and PyTorch ≥ 2.0.

---

🧩 Dataset Layout

Example folder structure:

/path/to/RICE_DATASET/ ├── train/ │ ├── cloud/ # hazy or cloudy input images │ └── label/ # clean ground-truth targets └── test/ ├── cloud/ └── label/

---

📊 Datasets

No.	Dataset	Year	Pub.	Number	Image Size	Type	Download
01	SateHaze1k	2017	WACV	400×3	512×512	Synthetic	🔗 Download
02	RICE (RICE1 & RICE2)	2019	arXiv	950	512×512	Synthetic	🔗 Download

Notes:

• SateHaze1k is divided into Thin, Moderate, and Thick subsets simulating different haze densities.
• RICE1 and RICE2 include paired cloudy and clear-sky remote sensing images for evaluating dehazing and cloud removal performance.

---

🧱 Step 1: Grid-Based Dataset Preparation

To generate 128×128 tiles for training and validation:

python GLADE-Net/tools/grid_crop_pairs.py \
  --root    dataset \ #path to your dataset train[cloud, label], test[cloud, label]
  --out_root dataset_grid \
  --tile_w 128 --tile_h 128 --overlap 0

Output tiles will be created under: dataset_grid/train/ dataset_grid/test/

---

🚀 Step 2: Training Phase 1

Run training from the project root:

cd GLADE-Net
python -m src.train \
  --train_dir dataset_grid/train \
  --val_dir   dataset_grid/test \
  --img_size 128 --batch_size 1 --epochs 2500 --devices 1 \
  --precision 16 --accumulate 2 --lr 2e-4 \
  --save_ckpt Phase_1.ckpt

You can resume training with:

python -m src.train --train_dir ... --val_dir ... --resume /path/to/checkpoint.ckpt

---

🔍 Step 3: Inference Phase 1 (Grid UFormer)

Run inference on cloudy test images using the trained checkpoint:

python GLADE-Net/tools/inference_grid.py \
  --weights Phase_1.ckpt \
  --input   dataset/test/cloud \
  --output  dataset/test/output_phase_1 \
  --tile 512

The script will:

• Load each image, pad to a multiple of the tile size.
• Partition into non-overlapping grids.
• Apply the UFormer model patch-by-patch.
• Reconstruct and crop back to the original size.
• Save restored outputs under output_phase_1/.

Example output folders: dataset/test/cloud/ # input hazy images dataset/test/label/ # ground truth dataset/test/output_phase_1/ # Phase 1 restored results

---

🧮 Step 4: Training Phase 2 – Perceptual GAN Refinement

The next phase will use the output from Phase 1 (restored dehazed images) paired with their ground-truth targets to train a perceptual enhancement GAN.
This stage focuses on:

• Fine texture recovery and color correction.
• Multi-domain learning (RGB + Laplacian).
• Perceptual and adversarial loss integration.

python GLADE-Net/Phase_2/code/train.py\
     -opt GLADE-Net/Phase_2/code/options/train/train_second_phase.json

---

🔍 Step 5: Inference Phase 2

Run inference on cloudy test images using the trained weight:

python GLADE-Net/Phase_2/code/test.py\
     -opt GLADE-Net/Phase_2/code/options/test/test_second_phase.json 

---

🚀 Quick Demo – Try GLADE-Net in One Notebook

If you want to quickly test GLADE-Net (Phase 1 + Phase 2) without setting up the entire training pipeline, we provide an interactive demo notebook.
You can open and run it directly on GitHub or in Google Colab.

Demo Notebook:

📘 This notebook demonstrates:

• Loading pretrained Phase 1 and Phase 2 models
• Running inference on sample remote-sensing images
• Visualizing haze removal and perceptual enhancement results

---

🧠 Model Weights

Pretrained model weights for GLADE-Net are available for both phases and all datasets.
You can download them from the shared Google Drive folder below.

🔗 GLADE-Net Weights (Google Drive)

Folder Structure:

Phase	Folder	Dataset	File Name	Description
1	pahse_1	SateHaze1k	SateHaze1k.ckpt	Global attention model trained on synthetic haze data.
		RICE-1	RICE_1.ckpt	Phase 1 model trained on RICE1 dataset.
		RICE-2	RICE_2.ckpt	Phase 1 model trained on RICE2 dataset.
2	phase_2	SateHaze1k	SateHaze1k.pth	Laplacian-Guided GAN model for perceptual refinement.
		RICE-1	Rice_1.pth	Phase 2 refinement model trained on RICE1 dataset.
		RICE-2	Rice_2.pth	Phase 2 refinement model trained on RICE2 dataset.

---

Notes:

• Each .ckpt file corresponds to the Phase 1 (UFormer) stage for global dehazing.
• Each .pth file corresponds to the Phase 2 (GAN Refinement) stage for perceptual enhancement.
• Models can be directly loaded using the provided demo notebook or inference scripts.

---

🧾 Output Results

The final qualitative and quantitative results of GLADE-Net are available for each dataset and test condition.
You can browse and download them from the shared Google Drive folder below.

🔗 GLADE-Net Results (Google Drive)

Folder Structure:

Folder	Description
Grid_SPSR_test_thin	Results on SateHaze1k–Thin (light haze conditions).
Grid_SPSR_test_Moderate	Results on SateHaze1k–Moderate (medium haze).
Grid_SPSR_test_thick	Results on SateHaze1k–Thick (dense haze).
RICE_results	Dehazing results on RICE1 and RICE2 datasets.
test real images hazing	Visual qualitative results on real-world hazy aerial images.

🧩 Each subfolder contains restored RGB images from the Phase 1 (Global Attention) and Phase 2 (Laplacian-Guided GAN) pipelines, organized according to haze intensity and dataset type.

---

🧠 Citation

If you use this repository, please cite the paper: Common soon

---