Sea Ice Segmentation : Convolutional neural Networks(CNN) - Gray Level Co-occurence Matrix(GLCM)

Classification of sea ice and water in SAR imagery using a CNN and K-Means clustering for pseudo-labeling. It extracts image windows and GLCM features, normalizes and quantizes images, and generates a prediction map for visualization. This approach enhances the analysis and understanding of ice and water regions in the Sentinel-1 Satellite Imagery.

1. User Authentication

• Purpose: Establishes a connection with NASA's data services.
• Details:
  • Utilizes the grimp.NASALogin class to create a login session.
  • Authenticates the user with a username and password, allowing access to remote datasets for subsequent data requests.

2. Data Retrieval

• Purpose: Fetches SAR imagery based on specific filters.
• Details:
  • We utilize Greenland Ice sheet Mapping Project (GrIMP)-data products hosted by Nasa Snow and Ice Data Center(NSIDC) to retrieve SAR images
  • Uses the grimp.cmrUrls class to search for images within specified time frames (e.g., 2019 and 2022) and product types (e.g., sigma0 or gamma0).
  • We obtain an nisar image series according to bounding box,product type and time range specified

3. Geospatial Processing

• Purpose: Clips the retrieved images to focus on specific geographical areas.
• Details:
  • Creates a bounding box using geometrical bounds defined in a GeoDataFrame .
  • Employs the subset and clip methods to restrict images to the area of interest, enhancing data relevance for further analysis.

4. Feature Extraction & Label Mapping - GLCM & KMeans Clustering

• Purpose: extracts meaningful features for analysis.
• Details:
  • Utilizes the extract_windows_and_features function to create smaller windows for detailed analysis, calculating texture features using the Gray-Level Co-occurrence Matrix (GLCM).
  • Extracts properties such as contrast, dissimilarity, homogeneity, ASM (Angular Second Moment), energy, and correlation.
  • Organizes extracted windows and features into arrays suitable for training machine learning models using the prepare_dataset function.
  • K-Means Clustering: The code concatenates image windows and GLCM features into a single dataset and applies K-Means clustering with two clusters. It assumes that cluster 0 represents water and cluster 1 represents sea ice, mapping the labels accordingly.

5. Model Definition & compilation

• Purpose: Constructs a convolutional neural network (CNN) model for classification tasks.
• Details:
  • Creates two input branches: one for image windows and another for extracted GLCM features.
  • The image branch consists of several convolutional layers, max pooling layers, and dropout layers, flattening into a dense layer.
  • The GLCM features branch contains dense layers to process the numerical feature set.
  • Outputs from both branches are concatenated and passed through additional dense layers, culminating in a final output layer with a sigmoid activation function for binary classification.
  • Utilized the Adam optimizer and binary cross-entropy loss function, suitable for binary classification tasks.