Studentinformatie

MSc thesis topic: Improving the robustness of Neural Networks against missing inputs

Many problems in Remote Sensing, including change detection and land-use mapping, require the use of temporal series of satellite imagery or can benefit from them.

Image classification models based on Machine Learning and, in particular, Deep Learning models such as Convolutional Neural Networks (CNNs), have shown impressive results in many problems involving Remote Sensing imagery, such as object detection and land-cover mapping. Most models require a fixed dimensional input, and are therefore not well adapted to cope with missing inputs, as can arise due to cloud cover.

We want to explore a possible way of improving the robustness to partially or totally occluded images with CNN-based methods applied to times series of remote sensing imagery. In particular, the project would consist on adapting Dropout, a widely used technique in CNNs. Dropout is meant to prevent internal CNN activation from depending too much on particular combinations of activations in the previous layer. We would like to modify this technique to work at the input level, such that if one input image is missing, the model adapts to give the best possible answer given the remaining inputs.

Objectives

  • Research the usefulness of using Dropout at the input level to improve the robustness of CNN-based methods to missing inputs.
  • Apply the developed method to the case multi temporal imagery with partial cloud cover.

Literature

  • Che Z, Purushotham S, Cho K, Sontag D, Liu Y. Recurrent neural networks for multivariate time series with missing values. arXiv preprint arXiv:1606.01865. 2016 Jun 6.
  • Volpi M, Tuia D. Dense semantic labeling of subdecimeter resolution images with convolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing. 2017 Feb;55(2):881-93.
  • Srivastava N, Hinton GE, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning research. 2014 Jan 1;15(1):1929-58.

Requirements

  • Programming in Matlab or Python (or high motivation for learning).
  • Some background in statistics or Machine Learning is a plus.

Theme: Modelling & visualisation