Cloud Segmentation with Titan and CloudSen12

This project focuses on the preparation and training of YOLO nano for cloud segmentation, using the CloudSen12 and Titan datasets.

---

Step 1: Dataset Preparation

CloudSen12 Dataset Preprocessing

Description

• Filters samples with cloud coverage > 55%
• Uses only the Sentinel-2 RGB bands
• Extracts polygons of cloud classes 1 and 2 (these classes represent clouds)
• Converts masks into YOLO-Segmentation format

How to Run

1. Install dependencies:

   pip install -r requirements.txt

2. Run the script:

   python preprocess_earth.py

3. Output: RGB images and .txt annotations in YOLO-Seg format saved in:

   datasets/CloudSen12/train/
   datasets/CloudSen12/val/
   

---

Titan Dataset Preprocessing

Description

• Annotated using LabelMe (polygons)
• Automatically merges and splits into train, val, and test
• Converts .json annotations into YOLO-Segmentation format
• Removes corrupted images or those without valid labels
• Merges train + val → full_train

How to Run

1. Make sure the original dataset from https://zenodo.org/records/13988492 is placed inside the datasets directory, and each subdirectory test and traincontain only labels and images, others subdirectory must be deleted.

2. Run the script:

   python preprocess_titan.py

3. Output: Data and annotations in YOLO-Seg format saved in datasets/Titan/

---

Step 2: Train and Tune the YOLO Model

Installation

Install Ultralytics YOLOv11:

pip install ultralytics

and install all dependencies via requirements.txt:

pip install -r requirements.txt

---

Training Workflow

Run the scripts in the following order:

1. training_earth.py – Initial training on CloudSen12
2. tuning_titan.py – Fine-tuning on the Titan dataset
3. final_model.py – Retrain the final model on the entire Titan dataset and evaluate its performance

Make sure the .yaml files in the yolo_configs/ folder point to the correct dataset paths.

---

Retuning or Retraining

To retrain or experiment with new parameters:

• Change the model names and weights in the scripts
• Update the .yaml files if needed

---

Project Structure

.
├── datasets/
│   ├── CloudSen12/
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   └── val/
│   │       ├── images/
│   │       └── labels/
│   ├── Dataset_Zenodo/
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   └── test/
│   │   │   ├── images/
│   │   │   └── labels/
│   └── Titan/
│       ├── full_train/
│       ├── train/
│       ├── val/
│       └── test/
├── scripts/
│   ├── final_model.py
│   ├── preprocess_earth.py
│   ├── preprocess_titan.py
│   ├── training_earth.py
│   └── tuning_titan.py
├── yolo_configs/
│   ├── earth.yaml
│   ├── titan.yaml
│   └── titan_full.yaml
├── requirements.txt
└── README.md

---

Notes

• To obtain a consistent subdataset from CloudSen12 in preprocessing phase we applied some controls in order to obtain valid images which can make the download time-consuming (depending on your internet speed). We therefore recommend using the pre-uploaded version.
• Make sure to run the scripts from the correct directory

Dataset

Yahn, Zachary; Trent, Douglas; Duncan, Ethan; Seignovert, Benoit; Santerre, John; Nixon, Conor. (2024).
Supplemental Data: Rapid Automated Mapping of Clouds on Titan with Instance Segmentation (2.1) [Data set]. Zenodo.
https://doi.org/10.5281/zenodo.13988492

Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).