Neon flying squid (Ommastrephes bartramii) is one of the economically important cephalopod species in the Northwest Pacific Ocean. Their spatial distribution and abundance are synergistically regulated by multiple factors under the marine environment in sustainable fisheries. However, conventional modelling is constrained by the inherent sparsity and zero-inflation problems of fishery catch data, leading to distortions in initial information representation. Moreover, these models cannot accurately capture the complex nonlinear relationships between environmental conditions and resource dynamics, thereby limiting high-precision prediction and mechanistic interpretation. In this study, a deep learning framework was developed to integrate multi-source environmental elements for simultaneous prediction and intelligent interpretation of squid resources. The inputs were taken from the multi-annual (2015–2019) satellite-derived environmental data and fishery catch records. The framework included the enhanced TimeXer prediction, the hierarchical classification, and the interpretability module. Among them, 1) the TimeXer prediction module was constructed to effectively capture complex interdependencies between multi-source environmental variables—including sea surface temperature (SST), chlorophyll-a concentration (Chl-a), salinity, dissolved oxygen (DO), pH, and sea surface height (SSH) at multiple depths (0, 100, 200, and 300 m)—and historical Catch Per Unit Effort (CPUE) sequences. An adaptive gating mechanism was further employed to dynamically fuse these heterogeneous information streams, significantly improving CPUE prediction accuracy. 2) The hierarchical classification module with LightGBM was implemented to mitigate data imbalance. A two-stage classification was performed: Firstly, the fishing areas were identified from non-fishing areas, followed by fine-grained categorization of resource abundance levels (Few, Little, Mid, and Most) within the region of Interest (ROI) fishing zones. In the interpretability module, the SHAP (SHapley Additive exPlanations) values were integrated to quantitatively assess the marginal contributions and interactions of each environmental factor. A transparent and systematic analysis of key drivers was obtained for the squid distribution patterns. Experimental results demonstrated that the superior performance of the framework was achieved on an independent test set, with an R² of 0.9722 and a Mean Squared Error (MSE) of 0.0308, significantly outperforming several state-of-the-art deep learning models (e.g., Transformer, Crossformer, and FEDformer) over multiple metrics (MSE, RMSE, MAE, and MRE). The primary environmental drivers were identified for the formation of high-catch areas after SHAP analysis, including the chlorophyll-a concentration (optimal range: 0.1–0.5 mg/m³), sea surface salinity (33.0‰–33.5‰), sea surface temperature (15.0–22.0 °C), dissolved oxygen, and sea surface height. Crucially, there were significant synergistic effects between vertical salinity stratification and latitude, chlorophyll-a and salinity, as well as salinity and dissolved oxygen. In addition, the complex "dynamic-nutrient-physicochemical" mechanism was attributed to the squid aggregation. This finding can provide an interpretable, high-precision framework to forecast the squid resource dynamics and decipher the synergistic driving mechanisms of multiple environmental factors. Some insights into the ecological habits of O. bartramiiunder can be obtained for the decision-making on the conservation and utilization of oceanic fishery resources.