Background: Variable mussel production
Greenshell™ mussel production is the largest aquaculture revenue earner in New Zealand. Of this, Pelorus Sound accounts for about 68% New Zealand’s production, where harvests are about 75,000 tonne green weight annually, generating around $200 M per year in revenue.
We know, however, that this production is not constant year-on-year. For example, records collected by Sealord Shellfisheries between 1997 and 2005 show a large decline and then upswing in meat yield anomaly over a six-year period, representing more than 20% variation in mussel meat production (Figure 1).
At the same time Sealord were collecting these records, the Marlborough Sounds Shellfish Quality Programme (MSQP, funded by the mussel industry and NIWA) were collecting water quality data throughout Pelorus Sound. Those data showed that the abundance of the food of mussels, seston, varied coincidentally with the mussel meat yield (Figure 1). Therefore, we concluded that variable food supply was at the root of the variation in mussel meat production.
Figure 1. Variation in mussel meat yield (solid line), and their food (seston: dashed line), between 1997 and 2005, as measured by Sealord Shellfisheries and the Marlborough Sound Shellfish Quality Programme (MSQP) programme. Data are expressed as anomalies from the long-term average.
What ‘drives’ the variation?
Further investigation showed that climate is a main driver of the variation in seston availability. We found that climate conditions occurring in the summer and winter halves of the year affected mussel yields differently. The pathways by which this operates are shown in Figure 2.
Ocean upwelling is a strong driver in the summer months (October - March) because it supplies dissolved nitrogen from the deep waters of Greater Cook Strait. As these enriched waters are drawn into Pelorus Sound by estuarine circulation, it stimulates seston production, increasing mussel yield. The upwelling is driven by westerly winds in the Greater Cook Strait region, most common in El Niño summers. In contrast, La Niña summers are dominated by easterly winds, resulting in less upwelling and lower seston concentrations. Therefore, the presence of El Niño and westerly winds are strong drivers of increased mussel yield in the summer months. NIWA has statistical models that enable us to forecast the future state of the El Niño/La Niña condition with moderate confidence three months into the future.
The story in the winter months (April - September) is somewhat different. At this time of year, seston concentration and mussel yield are most strongly associated with increases in Pelorus River flow. Wetter-than-average years have more nutrients added to the sound, increasing mussel yield. The freshwater also decreases the tendency of the water column to mix vertically, enabling phytoplankton to grow better in the well-illuminated, upper parts of the water column.
Figure 2. How climate affects mussel yield in Pelorus Sound. The left panel shows maps of sea surface temperatures (SST) derived from satellite data, in summer and winter across central New Zealand. Warm temperatures are shown by ‘warm’ colours and also higher values of the superimposed contour lines. In summer, relatively cool, upwelled waters are clearly seen in Greater Cook Strait including the area of Pelorus Sound entrance. In winter, upwelling disappears as sea temperatures cool everywhere, and waters are mixed by winter storms. The right panel describes the chain of climatic effects that drive seston concentration and mussel yield. In the summer half-year (October - March) effects are mainly from oceanic effects (El Niño, westerlies and upwelling), whereas in winter (April - September) the local effects of increased river flow become more important. See Zeldis et al. (2008) and Chiswell et al. (2016) for more details.
How we make the forecasts of mussel yield
To create a forecasting tool, we used statistical models to combine multiple climate variables with measures of historic mussel yield. We used a historical data set of yields measured on the harvesting vessels when mussel lines were harvested between 1997 and 2005 in Pelorus Sound. We used the mussel yield observations and the coincident climate values, to find the best combination of climate variables to predict those yields. This is detailed in Zeldis et al. (2013).
In the 2013 work, we used monthly values of the El Niño-Southern Oscillation (ENSO; a measure of the strength of El Niño/La Niña conditions), wind directions and strengths from wind monitoring stations in Cook Strait (Farewell Spit and Brothers Island), satellite-derived sea temperature at Pelorus entrance and gaugings of river flows measured at Bryant’s Stream in the Pelorus River. All of these data were available from international or NIWA databases for the 1997-2005 period. We then compared our predicted values of mussel yield with those actually observed, to gain understanding of the reliability by which we could predict yield from climate variables.
However, the next steps in terms of generating forecasts of yield from forecasts of climate, presented a problem which we had to overcome. We are not able to forecast monthly values of sea temperatures, winds or river flow at particular ocean locations, wind stations or river gauging stations as used in the original work (which used historical observations; Zeldis et al. (2013)). We therefore needed ‘proxy’ climate variables, which emulated those original variables but could also be forecasted.
To get a proxy variable for winds, we took advantage of the fact that wind directions and strengths are a function of air pressure differences across New Zealand (in this case, between Auckland and Christchurch). Because these air pressures can be forecasted, this enabled us to predict future wind conditions. For the river flow proxy, we estimated runoff into the Pelorus catchment using forecasted rainfall, evaporation and temperature, all of which can be forecasted. For ENSO, we used an index called Niño 3.4, which is the average sea temperature anomaly in the central Pacific Ocean.
We forecast all of the proxy variables above using a strategy known as ‘analogue forecasting’: essentially, using the past, to predict the future. New Zealand climate databases of past conditions (about 45 years long) are searched to find those which most closely match current conditions. The climate which occurred in the 3 months following that previous time, is taken as the forecast from the current state to 3 months into the future. See Mullan & Thompson (2006) for more details.
We then use these forecasts of the climate variables to make predictions of future mussel yield for the coming season using our original climate-mussel yield model. Our work showed that climate conditions related to mussel yields occurring about two months later, because it takes about that long for nutrient-rich water to mix into the sound, to create seston, and then for the mussels to respond with improved growth. This, combined with the fact that NIWA climate forecasts are often made ‘seasonally’ for the upcoming three month period, indicates that mussel yield forecasting could extend up to five months into the future for the summer months.
Due to uncertainties in the statistical models, we are unable to predict mussel yield adequately using the winter half-year climate data. We therefore do not present any forecasts for the period June through October, inclusive. We do present forecasts based on the summer half-year climate forecasts, but only evaluate them as being above average, average or below average, after considering the imprecision of the estimates.
Validating the mussel forecasts
We can test how well our model predicts mussel yield by comparing the forecast mussel yield with yields which were subsequently observed in mussels harvested in Pelorus Sound over the relevant calendar period. Mussels were collected from Beatrix Bay every second month between August 2015 to June 2019. They were processed at NIWA, with mussel yield assessed per individual in a sample of 26 mussels with a shell lengths greater than 80 mm. This work was done as part of a larger NIWA study to assess relationships between mussel body condition and environmental drivers.
We accurately predicted mussel yield 80% of the time, with the mean observed mussel yield falling within the 80% prediction interval in 12 of the 15 overlapping forecast periods (Figure 3). However, observed mussel yields were significantly lower than forecast yields between December 2015 and February 2016. We think that these differences were due to inaccuracies in our climate model forecasts.
The mussel forecasting model uses climate forecasts to make its predictions, meaning that inaccurate climate predictions will impact our ability to accurately predict mussel yield. The 2015-16 austral summer was a period of neutral ENSO signals, making it difficult to predict the direction and strength of the climate variables used in the mussel forecasting model. For example, our climate model predicted westerly winds during this period but the observed winds were mixed. Wind direction is known to be an important driver of mussel growth in Pelorus Sound (Zeldis, Hadfield et al. 2013), with a greater proportion of westerlies leading to more seston and greater mussel growth (Zeldis, Howard-Williams et al. 2008). Therefore, it is possible that the error in predicted winds led to a lower observed mussel yield than predicted.
Despite such uncertainties, and given that we accurately predicted mussel yield 80% of the time, we have reasonable confidence in our mussel yield forecasts. The results indicate the importance of collecting mussel yield observations from the region to enable continuing validation of the predictions for model improvement.
Figure 3. The relationship between forecast and observed mussel yields expressed as a percentage of the 1997 - 2005 monthly means. Forecast mussel yields (green, grey, orange) are plotted at the centre of their forecast period, with the error bars representing 80% prediction intervals. Observed mussel yields (blue) are plotted on the harvest date, with the solid circles and lines representing the mean yield and interdecile range (10th to 90th percentiles) for each harvest period, respectively. No forecasts were made for the winter months.