Pattern recognition in long‐term Sooty Shearwater data: applying machine learning to create a harvest index

Rakiura Māori (New Zealand's southernmost group of indigenous peoples) have harvested the chicks of burrow-nesting Sooty Shearwaters (Tītī; Puffinus griseus) for generations. As part of the harvest process, some families have maintained annual harvest diaries, some dating back to the 1950s. We used generalized boosted regression models, a machine-learning algorithm, to calculate a harvest index that takes into account factors that could impact the numbers of birds taken on any given hunt. For predicted vs. observed values, r2 was between 0.59 and 0.90 for the nanao (first half of the season, when chicks are harvested from burrows during the day) and 0.67 and 0.88 for the rama (second half of the season, during which chicks are harvested from the surface at night). Exploration of the controlling factors of the models revealed that “day of season” plays an important role in predicting daily harvest during the second half of the season (the rama). The nightly tally in the rama peaked approximately halfway through (10–15 days in), which is probably related to the timing of birds emerging from burrows to fledge. The models also suggested that data from the rama (when chicks are 100–120 days old) may be the most suitable for long-term monitoring of populations of Sooty Shearwaters due to consistencies in calculated harvest indices between diaries. Nanao harvest indices, although less consistent, showed patterns similar to those of the rama. When comparing these data to the harvest indices calculated by general linear models by Clucas and colleagues, we found that the agreement between both indices was r2 = 0.31 and r2 = 0.59 for the nanao and rama, respectively. The use of machine learning to correct for extraneous factors (e.g., hunting effort, skill level, or weather) and to create standardized measures could be applied to other systems such as fisheries or terrestrial resource management.