ABSTRACT
The present study used nighttime visible satellite images to identify the daily presence and absence of the fishing vessel aggregations, targeting the Japanese common squid (Todarodes pacificus) in the coastal waters of south-western Hokkaido, Japan. Here, statistical (generalized additive model (GAM) and generalized linear model (GLM)) and machine learning models (boosted regression tree (BRT)) were developed using a 3-year (2000–2002) presence/absence information from squid fishing aggregations and environmental variables (night-time sea surface temperature (SST), chlorophyll-a (Chl-a) concentration, K d(490) (diffuse attenuation coefficients of downwelling irradiance at 490 nm), and bathymetry). Our findings showed that BRT outperformed the regression-based models in predicting the potential squid fishing zones during the validation period (2003). Results from BRT indicated that potential fishing zones were closely associated with water depth. Both SST and Chl-a concentration were also found highly influential to squid occurrence, while K d(490), which is related to the water transparency, showed relatively less impact on the squid distribution. The spatial predictions using daily data from 2000 to 2003 revealed the gradual eastward movement of potential fishing zones between June and December, consistent with the pattern of squid fishing activities. Four experimental fishing surveys were further conducted to validate and improve our model predictions against experience-based fishing surveys. The results showed that the squid catches using our model predictions in 2012 substantially exceeded the average catches of experience-based fishing in 2011.
