I employed KERAS' pretrained MobileNet CNN and fine-tuned it for our particular flood detection (image classification) task.
NOTE: This code was inspired and modified from the following source: https://deeplizard.com/.
"Flooding" images were collected by authors of the paper named "Detecting floodwater on roadways from image data with handcrafted features and deep transfer learning*", available at "https://arxiv.org/pdf/1909.00125.pdf".
"No Flooding" images were collected from an automated google image search script.
The data can be downloaded from the following link: https://drive.google.com/file/d/1oT07-wVuYj_hByn4FqUtIj63Yg_fe7dP/view?usp=sharing
Fine-tuned flood detection model can be downloaded from the follwing link: https://drive.google.com/file/d/1mF8NmMClUbKXYJoDdW3OYHhwiSknI5hk/view?usp=sharing
Normal or No Flooding images were collected from google image search, there may be irrelevant images in this category because the images were downloaded using an automated script.
The trained model performed quite impressively and got an accuracy score of over 98% and it was only trained for about 5 minutes.