🧊 Mendenhall Glacier Retreat Analysis (2014 – 2023)

Author: Jinal Vyas (ASU ID 1233053785)

This project investigates glacial retreat patterns of the Mendenhall Glacier, Alaska, using 10 years of Landsat-8 OLI/TIRS Collection 2 Tier-1 data. Data were processed and analyzed using Google Earth Engine (GEE) for large-scale imagery access, and Python for visualization, segmentation, and quantitative change assessment.

---

📁 Repository Structure

📦 Glacier-Retreat-Analysis
├── glacierAnalysis_DataPreparationReport_JinalVyas.ipynb - Colab.pdf
├── glacierAnalysisModelFinalAnalysisAndDeployment_JinalVyas_1233053785.ipynb - Colab.pdf
├── animation_RGB_GlacierChange.gif
└── README.md

⚠️ Note: If the Colab notebooks do not open due to GEE authentication or rendering issues, please refer to the PDF versions above — they contain all code cells, outputs, plots, and visualizations.

---

🛰️ Project Overview

Objective

To quantify the decline in glacier area over a decade by:

• Building a time series of satellite images (2014 – 2023)
• Extracting spectral bands and generating RGB composites
• Performing binary segmentation to identify glacier regions
• Measuring changes in glacier coverage over time

---

⚙️ Data and Tools Used

Component	Description
Satellite Data	Landsat-8 OLI/TIRS Collection 2 (Level 2 Surface Reflectance, Tier-1)
Data Source	Google Earth Engine (GEE)
Bands Used	SR_B3 (Green), SR_B4 (Red), SR_B5 (NIR), SR_B6–7 (SWIR)
Software/Frameworks	Python (Colab), GEE API, Rasterio, GDAL, OpenCV, PyTorch, segmentation_models_pytorch
Model	UNet with ResNet34/ResNet50/EfficientNet-B4 backbones
Loss Function	Dice Loss
Evaluation Metric	Intersection over Union (IoU) — approaching benchmark standard
Visualization	Matplotlib, Seaborn, Geemap (GEE maps)

---

🔬 Workflow Summary

1. Data Acquisition & Pre-Processing

   • Landsat-8 data fetched via GEE for each month between 2014 – 2023.
   • Applied cloud & shadow masking using QA_PIXEL bands.
   • Generated median composites per month to minimize cloud artifacts.
   • Clipped imagery to Mendenhall Glacier region of interest (ROI).

2. Exploratory Data Analysis (EDA)

   • Visualized RGB (4-3-2) composites and computed correlation matrices across spectral bands.
   • Evaluated NDSI (Normalized Difference Snow Index) to isolate snow and ice regions.

3. Segmentation Model Training

   • Created glacier / non-glacier binary masks using OpenCV.
   • Built a custom dataset loader with torchvision transforms (224×224 resized inputs).
   • Trained UNet models with ResNet34, ResNet50, and EfficientNet-B4 encoders to compare performance.
   • Chose ResNet34 backbone as optimal based on IoU and Dice Loss scores.

4. Post-Processing & Results

   • Generated glacier masks for each year and measured glacier area in km².
   • Produced an animated GIF (animation.gif) visualizing RGB imagery changes over time.
   • The animation highlights a progressive reduction in glacier coverage, visually confirming retreat.

---

📊 Results Overview

• Best Model: UNet + ResNet34 (backbone)
• IoU: Close to benchmark standards for binary segmentation (~0.85 – 0.9)
• Observation: Glacier area has consistently decreased across the decade (2014–2023).
• Visualization: Year-wise RGB frames and animated GIF depict the spatial retreat pattern.

---

🖼️ Visualization Highlights

• RGB Animation: animation.gif shows glacier color and texture variation over 10 years, offering a realistic temporal progression.
• Segmentation Comparisons: Each frame includes the original RGB image, ground-truth mask, and predicted mask for clarity.
• Matplotlib Plots: Display year-wise area decline and IoU trend.

---

💡 Key Takeaways

• Segmentation > Thresholding: Even for binary masks, deep segmentation models capture context and spectral nuances that grayscale thresholds miss.
• Consistent Monitoring: Using GEE and UNet provides a scalable pipeline for tracking glacier retreat worldwide.
• Visualization Matters: Animations and realistic RGB renders improve public understanding of climate change impacts.

---

🧭 How to Run (Notebooks)

1. Open in Google Colab: Data Preparation Notebook Model and Deployment Notebook

2. Authenticate Google Earth Engine when prompted.

3. Ensure segmentation_models_pytorch, rasterio, geemap, and GDAL are installed.

4. Run each cell sequentially to reproduce the workflow.

---

📚 References

• USGS Landsat-8 Collection 2 Surface Reflectance Guide
• Google Earth Engine Python API Docs

---

Prepared by: Jinal Vyas Course: Glacier Change Analysis & Remote Sensing Applications Date: November 2024