Forecast for 2022 Iceberg Season off Newfoundland, Canada

Grant R. Bigg & Jennifer B. Ross

A combined control systems and machine learning approach has been used to forecast the 2022 iceberg severity off Newfoundland, Canada. The control systems model was developed in Bigg et al. (2019) and achieved an 80% skill level in predicting whether an iceberg season would be above or below the mean I48N value over the verification period of 1997-2016. The machine learning approach has been detailed in Ross et al. (2021) and predicts properties of the iceberg year (for example the rate of change) as well as total yearly severity. Both approaches use the same environmental variables, namely Labrador Sea Surface Temperature (LSST), the NAO and the Greenland Surface Mass Balance, while the machine learning approach also allows a measure of auto-regression through having knowledge of previous years’ value of the appropriate measure. The updated Greenland Surface Mass Balance forcing data is courtesy of Edward Hanna of the University of Lincoln.

Recap of recent forecasts

The WERR control systems model forecast has been released to the International Ice Patrol (IIP) every year since 2017. The machine learning addition has been included in part since 2020. The forecast for the 2017 season was 766±297, with an observed total of 1008, this being a success. In 2018 the forecast was 685±207, with an observed number of 208. While the forecast was lower than the 2017 number it was still too high, thus a failure. The 2019 forecast for August was for 516±150 icebergs. By the end of the season the total number had reached 1515. The 2019 forecast was therefore also a failure. The 2020 forecast (including the machine learning approach) was for low/medium ice year with 584±303 icebergs by the end of July. The observed number by this point was 221, therefore a low year, but approaching the threshold for medium which is set by the IIP at 230 icebergs. However the lower bound on the WERR forecast was 281, so was still slightly too high. Nevertheless, as the forecast was a combined WERR and machine learning approach, it was overall a success. The 2021 WERR forecast was for 675±123 icebergs, and the machine learning models agreed with this prediction of a medium iceberg year. Additionally, the machine learning approach forecast a low/medium rate of change and a single peak in iceberg numbers in April. In fact, 2021 proved to be an extremely low iceberg year, with one iceberg reaching south of 48°N in February. The 2021 forecast was therefore a failure.

WERR Control Systems Model

The monthly cumulative totals can be seen in Figure 1, along with the average iceberg numbers for the last 10 and 50 years. It can be seen that while the WERR forecast is for a medium iceberg year, it is on the lower end of medium, far below the 10 and 50 year averages. In previous forecasts, the 10 year average has been significantly higher than the 50 year average, however it can be seen from the figure that due to the quantity of low years seen in the last decade, the averages are currently very similar. The IIP record a low iceberg year as less than 231 icebergs south of 48°N, therefore 2013, 2018, 2020 and 2021 were all low years. The IIP threshold for a high iceberg year is more than 1036, and therefore in the last decade 2014, 2015 and 2019 were high years. It is worth noting that, despite not being included in the 10 year average, 2010 and 2011 were also low iceberg years. This follows current forecast patterns that suggest that yearly iceberg severity past 48°N is likely to overall decrease in the future, but with occasional high years.

Figure 1. Plot of the 2022 iceberg forecast, including the average number of icebergs past 48°N by August in the last 10 and 50 years.

Machine Learning Aspect

The three machine learning tools used are: Linear Discriminant analysis, a Linear Support Vector Machine algorithm (SVM) and a Quadratic SVM algorithm. The model accuracy, F1 score (a measure of incorrectly classified cases) and RMSE (root mean squared error) can be seen in Table 1.

Table 1. An overview of the machine learning model accuracy and errors (from Ross et al., 2021)

	I48N Annual Total			Rate of Change			Number of Peaks			Peak Month
	Accuracy (%)	F1	RMSE	Accuracy (%)	F1	RMSE	Accuracy (%)	F1	RMSE	Accuracy (%)	F1	RMSE
Linear Discriminant	54.6512	0.67	0.7924	41.8605	0.33	1	55.814	0.80	0.6647	26.7442	0.18	1.3118
Linear SVM	50	0.70	0.8627	37.2093	0.38	1.023	56.9767	0.85	0.6559	26.7442	0.24	1.2152
Quadratic SVM	50	0.75	0.9022	41.8605	0.50	1.0173	60.4651	0.87	0.6288	30.2326	0.42	1.3725
Mean Skill level	1.55			1.21			1.16			1.12

The forecast for each of these models for the four prediction aspects can be seen in Table 2. In table format, for I48N and RoC (rate of change) ‘0’ represents a low year, ‘1’ a medium year, and ‘2’ and high year. For the Peak Month prediction, there are 4 potential outputs, aligning with ‘0’ January-March, ‘1’ April, ‘2’ May and ‘3’ June or later. These months have been chosen because April or May is historically the peak month in iceberg numbers passing 48°N. For the number of peaks prediction there are only two options: ‘1’ peak or ‘2’ multiple peaks.

All three models agree that 2022 will show one peak in April. However, when the models disagree, it is important to look at previously observed combinations to make sure the forecast is possible (for example one forecast from the model predictions is [2,0,1,1], which has never been observed between 1931–2017). Also, Ross et al. (2021) found that the Linear Discriminant outputs have the most similar distribution to observed results, and therefore may in some cases take precedence over the other two models. However, while a [2,1,1,1] combination is possible, as [1,0,1,1] is the average prediction across the models, this is the machine learning forecast for 2022. Additionally, the WERR prediction shows a low rate of change across the iceberg season, and a low/medium I48N. However, it is worth noting that [1,1,1,1] is the more commonly observed combination.

Table 2. 2022 Machine Learning Model Predictions.

	I48N	RoC	Peak Month	No. of Peaks
Linear Discriminant	2	1	1	1
Linear SVM	1	0	1	1
Quadratic SVM	0	0	1	1

Thoughts and Conclusions

Overall, while the combined forecast is for a medium iceberg year, the WERR model suggests that total iceberg numbers will be on the lower side of a medium year. Additionally, the variability in the machine learning I48N prediction reflects the complexity of iceberg prediction in the Newfoundland region, however the selected forecast of [1,0,1,1] (a medium iceberg year, with a low rate of change, and one peak in April) aligns with the WERR prediction.

References

Bigg, G. R., Y. Zhao, E. Hanna, 2019, Forecasting the severity of the Newfoundland iceberg season using a control systems model, J. Operational Oceanogr., doi:10.1080/1755876X.2019.1632128.

Ross, J.B., Bigg, G.R., Zhao, Y. and Hanna, E., 2021. A Combined Control Systems and Machine Learning Approach to Forecasting Iceberg Flux off Newfoundland. Sustainability, 13(14), p.7705.