To improve the spatiotemporal granularity of fishing-ground prediction for Dosidicusgigasin the southeastern Pacific， we evaluated how multiple time scales and environmental-factor combinations influence the distribution of the center fishing ground. Using fishery and environmental datasets from 2012 to 2021， we considered five temporal windows （3 d， 6 d， 10 d， 15 d， 30 d） and seven input combinations derived from sea surface temperature （SST）， sea surface height （SSH）， sea surface salinity （SSS）， and photosynthetically active radiation （Par）. The application effect index of fishing ground （AEI_FG） was introduced to define the center fishing ground as the prediction target. Because PAR degraded performance at the 30 d scale， we applied a weight-optimization scheme to mitigate its adverse effect. We compared three models： U-Net， generalized additive model （GAM）， and artificial neural network （ANN）. Results show that U-Net consistently achieved the best performance across schemes and accurately captured the spatial structure of center fishing grounds. Prediction accuracy depended strongly on the environmental-factor configuration， with the SST and SSS combination performing best overall. These findings demonstrate that deep learning can effectively integrate multi-source oceanographic information to enhance prediction accuracy for fishing grounds， providing methodological support for multi-timescale forecasting and intelligent fishery information services.