Satellite measurements are crucial in researching and understanding of geophysical phenomena. Even though there are many satellites orbiting Earth, it is impossible to measure the whole surface at all times. Consequentially, satellite measurements have significant gaps in space-time. One of the variables that satellites measure is sea-level anomaly (SLA). It describes the deviation of the sea surface height at some point in space-time from the global average computed over several past years. We address the problem of missing satellite measurements prediction. State-of-the-art methods solve this problem with approaches based on convolutions and therefore require measurements to be on a regular grid. To achieve this, a preprocessing step is needed to discretize the measurements onto a regular grid. However, this also introduces imprecision into the input. In this thesis, we propose PointSLA, a method that uses point clouds for dense prediction of SLA satellite measurements and therefore does not require discretization of measurements. Experimental results show the proposed model achieving RMSE 5% worse than state-of-the-art method DIVAnd while enabling greater space and time flexibility of the input and not requiring measurement discretization
|
