Computer Science Developer in Cary, NC, United States

I was given the task of leveraging recent developments in machine learning to provide highly accurate change detection in very high resolution (around one meter spatial resolution) bi-temporal aerial imagery.

To allow using the object recognition capabilities of convolutional neural networks while avoiding high sample complexity, I took a semi-supervised approach. I trained a Variational Autoencoders (VAE), which optimizes an information-theoretic bound on latent information and allows training on a mixture of labeled and unlabeled data.

I also modified the VAE to allow categorical and temporally-dependent latent variables in addition to the usual independent continuous latent variables seen in VAEs. This means that the neural network could learn, for example, that it was more likely for vegetation to transform into a building than the other way around. Areas with poor lighting in one-time steps could then use information from different time steps to decide on the class label.

The quality of the predictions was highly satisfactory, considering the low labeling requirements. The paper was accepted to IGARSS 2017, where I presented the results and discussed potential future improvements to be made.