Explore the Deep learning Pipeline from data processing of the raw data to model ingestible data while exploring the usage of semantic segmentation methods and deep learning optimization techniques on geospatial data
- Used model: U-Net, originally developed for medical image segmentation, a fully convolutional neural network (FCN)
- Used model: Deep Lab V3, a more recent approach than the U-Net architecture (better performance)
- explore the result of the two models with a pretrained Resnet encoder
- explore the result of different methods in regularization, such as data augmentation, batch normalization, dropout
- test methods on different datasets