flood_detection_tb

Dataset

• There are 31 images (complete) in folders such as nebraska_20170108t002112.
• Images are further divided into images of size 256*256.
• The folders have 4 types of images vv, vh, flood_label and water_body_label.
• The complete image has padding that is empty this empty space can cause problems while training and must be dealt with.

Preprocessing

• Built two functions which combine vh and vv image to give a single image.
• Built a function(out) to combine flood_label and water_body_label to only give a image which shows only a flood reigion.
• Built a dropper function which drops the image which has significant padding.

Modeling

• Found a simple model on stackoverflow for image to image.
• Model first scales down the iamge and then scales up the image.

Traning

• Could not train on complete traning set due to memory restriction.
• Used only one of the two function to combine vh and vv image, used dropper and out function.
• Due to laptop limitations, batch_size was set to 1.
• Trained model is stored in h5, json file.

Future work scope

• The model is not trained on overall dataset so one can pick up from there and work on it.
• There are opportunities for further improvement in building a better model.