mwBTFreddy Dataset

​

​

​

​

Presenters: Evie Chapuma & Grey Mengezi �

20th November, 2024

1

Kuyesera AI Lab https://kailab.tech/projects�ailab@mubas.ac.mw

Background

• The xBD/xView challenge created a dataset for building detection and disaster damage classification but lacks African disaster data.
• Southern African buildings, including those in Malawi, have some characteristics that are specific to the region, therefore training a model using the xBD dataset for disasters in Africa could lead to some bias.
• Malawi lacks datasets for building detection and damage classification tasks despite the increasing frequency of natural disasters.
• Developing localized datasets is essential for accurate building detection and disaster response in the region.

​

2

Kuyesera AI Lab https://kailab.tech/projects

Context

Joplin, US Chilobwe, Malawi

3

Kuyesera AI Lab https://kailab.tech/projects

Methodology Overview

1. Generating Grid lines:

• In order to avoid the issue of overlapping images, we use gridlines to help us in saving pictures .
• We used a Python script and the coordinates for the areas of interest.
• The output of this activity are kml files

4

Kuyesera AI Lab https://kailab.tech/projects

Methodology Overview

​

2. Saving images from Google Earth Pro:

​

• We chose two dates, pre and post for capturing the images.
• We chose a specific resolution and zoom level for consistency
• Our output for this activity is a dataset of images equal to the number of grid boxes that we had.

​

5

Kuyesera AI Lab https://kailab.tech/projects

Methodology Overview

3. Georeferencing Images:

• We georeferenced the images using QGIS, utilizing the HCMGIS plugin alongside Google Satellite Hybrid.

Exclusion Criteria for Images:

• Images that did not feature any buildings were excluded from the dataset. The absence of buildings in these images is due to the area being obscured by Soche Mountain.
• The output of this stage is a dataset of georeferenced images consisting of two files per image; TIF image and GCP points file.

​

​

​

6

Kuyesera AI Lab https://kailab.tech/projects

Methodology Overview

4. Cropping of Images:

• We used python scripts to crop and to rename the cropped images
• The reason for cropping them is to remove other buildings that appear in the picture but are not inside our grid box of interest
• The main output for this stage are images which have been cropped and renamed

​

​

7

Kuyesera AI Lab https://kailab.tech/projects

Methodology Overview

5. Labelling of Images:

• We chose QGIS for its ability to work with shapefiles, allowing us to define and classify building features accurately sing the xBD methodology.
• We then run a python script to generate CSV files from the labels(shapefiles) which was then used to generate json files corresponding to each image using another script.

8

Kuyesera AI Lab https://kailab.tech/projects

Results

1. Images: Georeferenced Google Earth images of the focus area taken pre and post Cyclone Freddy
2. JSON files : Building labels for each image

​

​

9

Kuyesera AI Lab https://kailab.tech/projects

Challenges

​

• Lack of high resolution satellite data for Malawi
• Georeferencing is not 100% accurate
• Poor building structures

10

Kuyesera AI Lab https://kailab.tech/projects

Applications

• Object Detection: Identifying and locating buildings and other structures within the images.
• Damage Classification: Assessing and classifying damage to buildings.
• Environmental Visualization: Visualizing flood damage across the entire landscape.
• Urban Planning: Supporting decisions on relocation, the development of drainage systems, and other infrastructure improvements.

​

11

Kuyesera AI Lab https://kailab.tech/projects

Thank you

12

Kuyesera AI Lab https://kailab.tech/projects�ailab@mubas.ac.mw