Towards Automatic Functional Model Specification for Distributed Scatterers Using T-SNE

The Dutch peatlands are a notoriously difficult region to monitor using InSAR. Low temporal coherence and signal-to-clutter levels necessitate the extraction of collective behaviour by the suppression of noise and clutter. Conventional techniques used to accomplish this include multilooking and phase-linking. The t-distributed Stochastic Neighbour Embedding (t-SNE) algorithm is a dimensionality reduction technique that aids in the analysis of large datasets. In this paper, we present an initial investigation into the suitability of the t-SNE algorithm to take the idea of extracting collective behaviour further. Similarly-behaved patches of land are automatically grouped together by the algorithm which aids in the specification of a functional model for that group. Our initial results show that the algorithm is able to successfully identify and group together areas in a scene which display similar behaviour over time. We also find that groups which display the same behaviour may also contain the same kinds of processing errors (for example unwrapping errors or cycle slips) and that these can also be automatically detected by the algorithm. We present this result as the first building block in an approach to smart InSAR data analysis which can learn from the data it is processing.