Smart SAR

Project

Smart SAR uses machine learning techniques to focus radar data.

• Focusing SAR data requires geometric/physical information or auxiliary (aux) data (velocity, pitch, roll, and yaw)
• Using machine learning, we can estimate the aux data and focus parameters to focus without any recorded aux data

System

Methods

Create a hierarchical machine learning environment by:

• Implementing child models with a fully convolutional neural network
• Training a child model to handle every individual focusing algorithm
• Implementing a parent model using the OpenAI Gym reinforcement learning environment
• Training the parent model to analyze input data and determine what focuses to apply (what child models to use) and in what order to apply the focuses
• Connecting the child models to the parent model to completely focus the image

Conclusion

Smart SAR does well at focusing with a model overtrained with highly correlated data (1), but fails for when trained using uncorrelated data (2):

Future work:

• Research and implement physics-based approach to estimate the geometry of the transmitter, receiver, and target
• Synthesize data under new k-spaces using the physics model

Learned:

• How to implement machine learning models using TensorFlow and Keras
• SAR phase is statistically random, so we need to focus on the geometry and not the image data itself