Fig. 6: Calibration curves for IceNet and SEAS5.
From: Seasonal Arctic sea ice forecasting with probabilistic deep learning
Observed frequencies of ice (sea ice concentration >15%) are plotted against binned probabilities of the sea ice probability (SIP). IceNet’s calibration curve is close to the ideal black dashed line, indicating that IceNet’s output probabilities are almost equal to the true likelihood of sea ice occurring. In contrast, SEAS5 overestimates sea ice probability. The calibration curves are computed over the test years 2018–2020 and all six lead times.
