Spatio-temporal Forecast of Fishing Density using U-Net

Spatio-temporal Forecast of Fishing Density using U-Net
Jeon, Ho-Kun; Cho, Hong Yeon; Lee, Chol Young; Park, Yong Gil
KIOST Author(s)
Cho, Hong Yeon(조홍연)Lee, Chol Young(이철용)Park, Yong Gil(박용길)
Alternative Author(s)
전호군; 조홍연; 이철용; 박용길
Publication Year
Time-series forecast models for a specific stationary point can be made using the time-series models such as ARIMA and LSTM. However, for a spatio-temporal forecast for a specific area, it is necessary to obtain time-series data per point and to build a spatio-temporal model considering the correlation or the continuity between the points. This study uses the deep learning network U-Net to predict the fishing density in space and time domains. Firstly, the state of fishing is identified by the speed of V-Pass data representing the location of the time-series fishing vessel, and then daily 0.01° gridded fishing density data is generated. Before constructing the model, time-series variation and significant periods are derived using the median coefficient of variation(medianCV) and periodogram per grid. The U-Net is designed to return six consecutive daily prediction grids after receiving 30 consecutive daily grid data. Among four years of grid data, learning, verification, and testing are assigned to 2 years (2018-2019, 694 days), one year (2020, 330 days), and one year (2021, 330 days), respectively. The model's performance was evaluated using the coefficient of determination(R2) and degree of agreement(IOA) for spatial and temporal forecasts. It is reveald that the medianCV and the IOA have an inverse relationship, meaning that the forecast is stable in the low-variation grid. The spatial performance has R2 0.87 in the forecast after 1 to 6 days from the last date of the given dataset.
Bibliographic Citation
International Conference on Aquatic Science & Technology (i-CoAST) 2023, pp.47, 2023
international Conference on Aquatic Science & Technology (i-Coast)
Files in This Item:
There are no files associated with this item.


Items in ScienceWatch@KIOST are protected by copyright, with all rights reserved, unless otherwise indicated.