This repository focuses on training semantic segmentation models to predict the presence of floodwater for disaster prevention. Models were trained using SageMaker and Colab.
Create an ensemble of the segmentation models. Use the max value of all the models as the ensemble value. This will bias the model towards saying it is floodwater rather than missing out on floodwater. This is great because in production we'd be less like to miss out on floodwater events (which is much worse than underpredicting them), and it also has been shown in other projects that the models will typically miss out on floodwater more often than not so this can correct for that.
To improve model performance, train some new models with data from Microsoft's Planetary Computer. This will augment the data with such things as elevation levels, making it easier for the model to detect floodwater.
This will be the final part of this project (unless I decide to make a future iteration). If I don't run into too many Windows deep learning issues, I will be doing a multi-gpu hyperparameter sweep with my local machine as well as on Colab.