Modelling Spat Performace

• Explore relationships between historical spat performance and environmental data using multiple regressions.

• Develop a decision-support tool to help farmers adapt spat-selection and seeding practices to improve crop performance.

• Assess if local environmental parameters influence spat performance.

• Use outcomes to guide future monitoring efforts and forecasting in AQU2023-05.

Industry data

• Industry partners supplied data in a variety of formats

• Sanford datasets already included PI ratios; Cedenco, NIML and Maclab required calculation

• For Cedenco and Maclab, seeding and interseeding records had to be matched by line and location to compute PI ratios

• Some partners supplied extra information (e.g. spat size), but these were not consistent across datasets

• Data focus on the early spat grow‑out period; where no interseeding occurred.

• Cedenco includes both primary‑to‑interseed and primary‑to‑late‑harvest records (i.e. no interseeding)

• All interseed stages data were removed

Industry data summaries

Sanford: 1,400 obs. from 91 sites in Banks Peninsula, TOS, SI (July 2007 - April 2024)

NIML: 939 obs. from 9 sites in Coromandel between (May 2013 - June 2023)

Cedenco: 439 obs. from 15 sites in TOS (March 2018 and August 2023)

Maclab: 199 obs. from 36 sites in TOS (October 2017 - November 2024)

Map of mean PI ratios

Response variable

P:I ratio is the meters of intermediate seed divided by the metres of primary seed, corrected for blue mussel numbers.

\[ P : I \text{ ratio} = \frac{\text{Intermediated seed (m)}}{\text{Primary seed (m)}} \]

Operational variables

• Seeded Date: date lines were seeded with fresh spat

• Strip Date: date lines were stripped (interseeding or harvest date)

• SiteID: farm identifier for primary seeding

• Owner: site owner (for confidentiality purposes)

• Region: broad grow‑out area

Cycle Time

Days between Seeded Date and Strip Date

Environmental data

High resolution hindcast data

• National model: NZ Environmental Datacube (Oceanum, for MPI & PFR); public data (2010–2023); ~300–400 m nearshore resolution; coarsest of the three.

• Marlborough model: 10-year hindcast (2009–2017) for Cook Strait/TOS (Develped by Cawtthron).

• Southland model: 10-year hindcast (2010–2019) for Southland–Stewart Island, developed for MPI.

• Variables extracted at 2 m intervals from 0–20 m depth: wind speed, wave height, current speed, salinity, temperature.

Modelling approach

Modelling process

• Removed irrelevant and collinear variables.
• Applied preprocessing (centering, scaling, transformation).
• Fit candidate algorithms (e.g. Boosted Regression Tree, Random Forest, Generalised Additive Modelling)
• Tuned hyperparameters automatically.
• Assess performance based on 5-fold cross-validation.

Predictor correlations

Stochastic Gradient Boosting

Partial dependence plots

CV observed vs predicted PIratio