ebgoldstein / mlstormimpacts

Classifier for detecting impacts in post-storm images

ML Washover

This is the dev area for creating a classifier for detecting washover in post-storm images. The cleaned version will live here.

Data

• Post-storm images from NOAA (Downloaded via psi-collect)
• Classifications from Hurricane Florence (data here)
• Extra data from Hurricane Michael (Testing)

Code

Directory Structure:

Should look like:

Users/ebgoldstein/MLWashover
                    ├── readme.md
                    ├── license
                    ├── data
                    │   ├── raw
                    │   ├── training                  
                    │   │    ├── nowashover
                    │   │    │     ├── C26047788.jpg
                    │   │    │     └── ...
                    │   │    └── washover
                    │   ├── validation
                    │   └── testing
                    └── src
                        ├── PictureSplitter.ipynb
                        ├── Mobilenet.ipynb
                        ├── Inception.ipynb
                        └── ...
                        

mlstormimpacts's Issues

grad-CAM

the implementation is here

need to visualize predictions from test set, then do grad-CAM overlays

flow preprocessing

flow the data from the figshare repository throught the preprocessing scripts.

transfer learning

w/ inception

class imbalance

https://www.tensorflow.org/tutorials/structured_data/imbalanced_data

Move to DAISY

when its done

transfer learning

w/ inception

confusion matrix is reversed

it is currently:

          False True
False
True

needs to be changed