PREDICTION OF THE CATCH OF JAPANESE SARDINE LARVAE IN SAGAMI BAY USING A NEURAL-NETWORK

We attempted to forecast the catch of post-larval stages of Japanese sardine with total length 19-35 mm exploited each year by troll and beach seine fisheries in Sagami Bay, Japan, during March and April. In the forecasting system, the feed forward (layered) type of neural network was utilized. The system for forecasting the catch in Sagami Bay during March and April was developed on the basis of (a) predicted hydrographic conditions (occurrence rates of the Kuroshio path types and distance between the axis and Cape Iroh-zaki in March and April) as predicted in the previous paper, (b) hydrographic data from November (previous year) to February (current year) and (c) Japanese sardine catch data in various landing regions in the previous year. The predicted values of catches agreed well with the observed catches. Upon investigation of the weights and threshold values in the trained neural network, the distance between the Kuroshio axis and Cape Iroh-zaki was found to significantly affect the predictions. We also examined how the data length for learning of the neural net affects the prediction. It appears that the neural network is a practical tool for predicting the catch.