Apply Deep Learning for image recognition techniques to a real world data set.
The original paper can be found here
The key objective of the project is to work on an open dataset consisting of satellite imagery data to detect damaged buildings due to a hurricane. These satellite images are taken from the aftermath of Hurricane Harvey in 2017. The goal would be to build a classifier which can identify if a specific region (based on the input satellite image) is likely to have suffered from flooding and other typical structural damages due to Hurricane Harvey.
- Do split the train dataset into 80:20 (train-validation split – use seed=42).
- Resize images to 128x128
- Build a basic CNN model from scratch (2-3) layers which should serve as a baseline model and give you around close to 80%+ accuracy easily
- In order to improve model performance, try
- Different approaches of data augmentation,
- more complex CNN architectures,
- pre-trained models,
- transfer learning along with NN training approaches like early stopping, dynamic learning rates etc.
- Showcase model performance on test data using confusion matrix, classification reports
- Deep Learning
- Data Visualization
- Predictive Modeling
- Python
- Pandas, Numpy, Scipy
- Sklearn,
- Tensorflow
- Clone this repo (for help see this tutorial).
- Raw Data is not included in this repo.
- Data processing/transformation scripts are being kept in
notebooks\