Feed-Forward Neural Network for ERS-2 Retracker Threshold Computation

Current release candidate v2.0 (RC2) of the ERS-2 retracker threshold model used for producing the satellite-altimetry-based sea-ice thickness climate data record (CDR) v3.0 of the European Space Agencies (ESA) Climate Change Initiative+ (CCI+) on sea ice.

Model training is based on orbit trajectory matches between ENVISAT and ERS-2 within the mission overlap period between 2002/10 and 2003/04. All training data was generated from trajectory matches within the Arctic basin and within a radius of 1.5 km around the each individual ERS-2 waveform. Initial optimal-retracker thresholds were then computed from ENVISAT average reference freeboards per ERS-2 waveform.

Model input are individual echo-waveform subsets of 35 range bins around the first-maximum index used the by the Threshold First Maximum Retracker Algorithm (TFMRA; 5 bins before and 30 bins after the first-maximum index) – threshold computations are therefore independent on any auxiliary data or associated waveform parameters.

Model architecture (pyTorch implementation):

fnn_envisat_rc1 (
  (fc1): Linear(in_features=45, out_features=2048, bias=True)
  (fc2): Linear(in_features=2048, out_features=2048, bias=True)
  (fc3): Linear(in_features=2048, out_features=2048, bias=True)
  (fc4): Linear(in_features=2048, out_features=2048, bias=True)
  (fc5): Linear(in_features=2048, out_features=2048, bias=True)
  (fc6): Linear(in_features=2048, out_features=1, bias=True)
)