Abstract: Abstract: To improve the accuracy and practicability of fishery forecast in the Northwest Pacific, a method of constructing a forecast model of squid was proposed based on the principle of deep learning. In this study, the data included the fishery catch data from the North Pacific squid fishing boat production information and the Sea Surface Temperature (SST) from the moderate-resolution imaging spectroradiometer, from July to November 2000-2015. According to the combination of different channels, four kinds of datasets were formed for the model training, including the single-channel dataset only containing SST; 2-channels dataset of SST and month; 3-channels dataset of SST, longitude, and latitude; 4-channels dataset of SST, month, longitude, and latitude. To match the data of the first channel in dimensionality, the three-input data of longitude, latitude, and month needed to be expanded from a 0-dimensional scalar quantity to a 2-dimensional tensor with pixels of 65×65 and regarded as the second, third, and fourth channel. Because of the insufficiency of effective fishery catch data, these datasets were enhanced by random rotation of the SST image with a small-angle between -10° and +10° and a random 0.1° offset of the image center in four directions, including north, south, east and west. The AlexNet was chosen as the structure of the Convolutional Neural Network (CNN) model, and it consisted of five convolutional layers, three max-pooling layers, and three fully-connected layers with a final 2-way softmax. Different from traditional fishery forecast methods, this method used the Graphics Processing Unit (GPU) to accelerate training, and its extraction of environmental features was automatically completed by computer. SST, latitude, longitude, and month were all factors that needed to be considered when constructing a fishing ground forecast model. The impact of these factors on the accuracy of the fishing ground forecast was compared and analyzed. The results showed that 1) According to the migration laws of squid, the datasets from July to November were divided into three sub-datasets, including July to August, September, and October to November. This way of month combination increased the testing accuracy by at least 6.1 percent points. The testing accuracies of three sub-datasets of July to August, September, October to November were much higher than that of the whole dataset (74.4%) from July to November. 2) The training result of the 4-channels dataset was the best, and the testing accuracy was significantly higher than that of others. The single-channel dataset only containing SST achieved the testing accuracy of at least 73.5%, which indicated that SST was the most important factor among the four factors of SST, longitude, latitude, and month. 3) The actual fishery catch data of 2015 was used to validate the accuracy of the forecast model, and precision and recall were chosen as the evaluation indexes of this model. The average precision, recall, and F1-score were 66.6%, 82.3%, and 73.1%, respectively. The predicted high-yield fishing areas basically matched the actual high-CPUE (Catch Per Unit Effort) areas, and the monthly movement trends of both were also basically consistent. 4) The training results were satisfactory, and the testing accuracy converged to about 80.5% after 80 000 iterations of training. The accuracy of three testing datasets with 4-channels dataset of July to August, September, and October to November was 80.5%, 81.5%, and 81.4%, respectively. It could be concluded that SST and its temporal and spatial information played an important role in the forecast of the Northwest Pacific squid fishery. And the training results demonstrated that it was feasible to construct a squid fishery forecast model by using a dataset of single environmental factor SST and CNN. It also could be concluded that the migratory laws of squid were significant and could not be ignored in the process of the fishery forecast model construction.