The dynamic monitoring of fishing activities is fundamental to fishery management. Leveraging multi-year (2020–2023) AIS data from squid jigging vessels, this study employed a multi-level data mining and spatial statistical approach to decode the spatiotemporal patterns of fishing effort in the Southeast Pacific Ocean. Our analysis reveals a highly concentrated and cyclical operation model: temporally, 20% of days contributed 46% of the total effort; spatially, 30% of the fishing grounds accounted for 80% of the effort, forming four persistent hotspots. Vessels exhibited a distinct bimodal speed distribution, enabling clear behavioral differentiation. While no fishing was detected inside the seasonal no-take zone, activities were observed near its boundaries and Exclusive Economic Zones, highlighting compliance and potential risks. The significant spatial aggregation, strongest in June, underscores the tight linkage between fleet operations and resource distribution. These findings provide a scientific basis for spatially explicit management strategies to ensure the sustainable harvesting of squid resources.

To accurately predict the fishing intensity of the Northwest Pacific squid fishing grounds and address the limitations of traditional models in capturing long-term temporal and spatial correlations and neglecting the coupling relationships of deep environmental factors, this study constructs a 3DCNN model and three fusion models incorporating residual, attention, and Transformer mechanisms. Using the 2017–2024 AIS fishing data and ocean environmental variables from the North Pacific squid fishing industry, the models’ performance is compared at 12 different temporal and spatial scales, and key core environmental variables are identified. The results show that the ResNet3D model exhibits the best overall performance, achieving an F1 score of 0.7909 at the 1.0°-7 days temporal–spatial scale. The residual connections effectively mitigate the gradient vanishing problem, balancing prediction accuracy and stability. The optimal spatial resolution is 1.0°, and the key environmental variables include S₁₀₀, Chl-a₁₀₀, PP₁₀₀, and DO₁₀₀. S₁₀₀ is the core driving variable, consistently exhibiting the highest feature importance value at all time scales. It should be noted that Chl-a is considered an indirect indicator of primary productivity, which may influence squid distribution through trophic transfer processes rather than direct biological effects. This study demonstrates the prediction accuracy and applicability of the multi-mechanism fusion 3DCNN model, reveals the temporal and spatial distribution patterns of fishing intensity in the Northwest Pacific squid fishing grounds, and provides scientific methods and technical support for dynamic monitoring, intelligent management, and sustainable utilization of squid resources.