| Summary: | Climate modelling and high-latitude marine navigation require improved information on sea-ice floe extents and dynamics. New satellite sensors provide raw data of this nature but the volume of information makes human analysis impractical. To address this problem, a software system for automatic tracking of sea-ice floes in satellite imagery has been designed and evaluated. Using a recurrent neural network model, experiments were conducted to discover suitable design parameters. Simulated imagery time-sequences of increasing complexity were produced to train successive models. The networks produced were evaluated based on performance and reliability. A small-scale working system, able to map multiple input features in image sequences to Cartesian coordinates, was produced. Results show that a recurrent neural network is suitable for the tracking task and has advantages in robustness and speed over other approaches. Recurrency (feedback) was found to be crucial in achieving good performance.
|
|---|