A computer vision approach to detect flooded areas from aerial imagery.
The aim is to detect areas with flooding and generate a 'patchwork quilt' showing flooded areas to enable better route planning for emergency services.
The dataset for this project can be found here:
This was created using the file slice_and_crop.ipynb in make_datasets dir. The source data was collected using the search term 'flood imagery' : http://qldspatial.information.qld.gov.au/catalogue/custom/search.page
This yielded these 2 datasets and more:
- Flood imagery - Talwood 0.5 metre - January 2011 2.Flood imagery - Surat 2 metre - January 2011 You can download my dataset by making an issues and I will email it to you :)
In training.ipynb exists code to training a simple CNN. Performance would likely improve with models such as NASNet or a more simple architecture such as Inception-v3.
I decided to use Keras since it's becoming the Tensorflow default API in TF 2.0! 🎼🔈 Times they are a'changing 🎼🔈 Run this in google colab! I've tested it to make sure it works! Just:
git clone https://github.com/francotheengineer/Flood_detector.git
and upload the ipynb to Colab!
The idea is to make a colour modified map to show the areas with and without flooding. I used another image from the QLD Spacial site for inference.
The output can be seen:
Another result before I performed some dataset cleansing:
-
There is a significant colour variation in the colour of the floods. This is the main artefact use to make the dataset. This can be solved with much more training data. <500 images isn't a lot for a full train of a CNN.
-
Some data normalisation could be used:
determine flood water colour range -> replace with a narrower range of colour -> do for both train and inference images -
Test more modern CNNs such as Inception-V3 and compare results
-
I spent quite some time dealing with the enormous images from the source. This was impossible to deal with in OpenCV due to 2^32 pixel limits. I resized with GIMP then read into OpenCV, change your slicing window accordingly! This flag can also work when the images are lower than the limit but use too much ram:
org_image = cv2.imread(image_path, cv2.IMREAD_REDUCED_COLOR_8)