Abstract

Climate change has profoundly impacted glacier ecosystems, leading to challenges such as glacier retreat and lake formation. Glacier Guard is a deep learning-based object detection system designed to monitor these changes using the YOLOv11 model. This project leverages a custom dataset comprising 208 satellite images of the Badswat Glacier, collected from Google Earth Pro, spanning 2013 to 2024. Data augmentation techniques expanded the dataset to 580 images, enhancing model robustness. The YOLOv11 model was trained to detect glacier retreat, lake formation, and water runoff, achieving promising results with high precision and recall. This report highlights the potential of integrating deep learning with satellite imagery for early warning systems and glacier monitoring. Future work aims to scale the project to more glaciers in Gilgit-Baltistan and enable real-time deployment.

Methodology

1 Data Collection
Satellite imagery of the Badswat Glacier was acquired using Google Earth Pro, capturing temporal changes from 2013 to 2024. The curated dataset includes 208 images, highlighting features such as lake formation, retreat areas, and water runoff.

2 Data Augmentation
To improve model generalization, the following augmentation techniques were applied:

Flipping: Horizontal and vertical flips.
Rotation: Random rotations (±15°).
Saturation Adjustments: Variations in saturation to simulate diverse lighting conditions.
The dataset was expanded to 580 images, creating a robust foundation for training.

3 Model Architecture
The YOLOv11 object detection model was chosen for its state-of-the-art accuracy and real-time performance, making it well-suited for satellite imagery analysis.

4 Training
Hyperparameters used during training:

Epochs: 50
Batch Size: 16
Image Size: 224x224
Learning Rate: 0.01
The model was trained to detect three classes:

Lake Formation
Retreat Areas
Water Runoff

5 Validation
A validation set of 21 images was used to evaluate model performance. Metrics included precision, recall, mean Average Precision (mAP), and F1 score.

Results

The YOLOv11 model achieved the following performance metrics:

Precision: 0.903
Recall: 0.781
mAP: 0.83
F1 Score: 0.838

The model demonstrated high accuracy in detecting lake formations, retreat areas, and water runoff, with particularly strong performance in identifying lake formations.