Dynamic label and ConvLSTM-enhanced U-Net for Antarctic krill distribution prediction

Antarctic krill ( Euphausia superba ) occupies a pivotal niche in the Southern Ocean ecosystem and is also an important target species for commercial fishing, yet its distribution remains highly susceptible to climate change and environmental variability. Precise prediction of krill distribution is critical for achieving efficient harvesting and ensuring sustainable resource management. However, traditional static prediction models often struggle to capture the complex spatiotemporal evolution inherent in polar environments and are frequently constrained by the spatiotemporal discontinuity of observational data. To address these challenges, this study proposed a U-Net-LSTM model to predict potential core habitats of krill. We constructed a dynamic labeling strategy based on Kernel Density Estimation (KDE) to generate robust binary targets (core/non-core habitats) from biological density records for training prediction models. Results indicated that the U-Net-LSTM model significantly outperformed traditional machine learning baselines (RF, GAM, KNN) and the standard U-Net, improving Recall and F1-scores by an average of 52.5% and 28.4%, respectively. Furthermore, the coverage rate of valid observation points in the test set increased by 23.3% compared to the static U-Net and by a substantial 76.8% compared to traditional machine learning models, demonstrating the model's superior capability in resolving the complex spatiotemporal distribution characteristics of E. superba . Crucially, sensitivity analysis of environmental lag effects revealed an optimal input sequence length of 7 months ( T = 7), which ecologically corresponds to the critical phenological transition from the sea-ice-dependent overwintering phase (July–September) to the peak summer feeding season (January–March). These findings provide significant methodological advances for E. superba habitat modeling and offer scientific support for the dynamic management of Antarctic fisheries under climate variability. • A KDE-based dynamic labeling strategy is proposed to generate robust core/non-core habitat labels, mitigating biases from sparse polar survey data and phenological shifts effectively. • The U-Net-LSTM model outperforms traditional ML (RF, GAM, KNN) and static U-Net by 52.5% (Recall) and 28.4% (F1-score) on average, capturing complex spatiotemporal dynamics of Antarctic krill. • Optimal input sequence length ( T = 7 months) of the U-Net-LSTM model corresponds to krill's transition from sea-ice-dependent overwintering to summer feeding, capturing critical cumulative environmental drivers. • Gradient-based attribution reveals key drivers (SID > CHLA > SST) and their spatiotemporal lag effects, unlocking the deep learning model's ecological decision logic.