AIM

  • Create a dataset of images taken from Google Earth Pro from selected areas in Blantyre affected by Cyclone Fredd. Each image is taken at two distinct dates, before and after Cyclone Freddy.
  • Categorize and label buildings in these images with two classes: damaged and not damaged.

OBJECTIVES

  1. Collect and prepare data for analysis.
    • Collect a dataset of images (pre-and-post disaster) from satellite sources.
    • Review the collected dataset to select a specific focus area.
  2. Apply the xBD methodology
    • Understanding of the xBD methodology for assessing building damage.
    • Apply the xBD methodology to the selected dataset of satellite imagery.
    • Produce an annotated dataset of images for the specified focus area based on the xBD methodology.
  3. Train the model and evaluate
    • Train a model using the xBD methodology on the annotated dataset of satellite imagery.
    • Test the trained model on other datasets to evaluate its performance.

HIGH-LEVEL METHODOLOGY

drm_sample_image

We use Google Earth Pro to obtain high resolution images of the selected areas at two different dates, before and after cyclone Freddy. Cyclone Freddy happened on the 11th and 12th March, 2023. To capture the images in Google Earth Pro, gridlines are added using a custom code to generate grids with specific latitude and longitude spacing, ensuring images don't overlap. Images are captured before and after Cyclone Freddy, with pre-disaster images from August 30, 2022, and post-disaster images from June 14, 2023, at 1024x768 resolution and 70m scale. These images are labeled by grid boxes and georeferenced in QGIS using four corner points of each box. The georeferenced images are then cropped and resized to 1024x1024 using a Python script, removing irrelevant areas and borders, and renamed systematically. Finally, the images are labeled in QGIS following the xBD methodology to classify building damage, generating CSV files and using a script, we generate JSON files for each image to create a comprehensive dataset of labeled, georeferenced, and cropped images.

EXPECTED RESULTS

  • Comprehensive image dataset from satellite sources (pre and post disaster)
  • Application of the xBD Methodology to the dataset, resulting in an annotated image dataset
  • A trained model capable of assessing building damage using satellite imagery.
  • Insights for Disaster Management - strengths and limitations of using satellite imagery for disaster management

TEAM

Dr. Amelia Taylor

Lewis Msasa

Evie Chapuma

Grey Mengezi

Chimwemwe Chamangwana