Time-prior-based stacking ensemble deep learning model for ship infrared automatic target recognition in complex maritime scenarios

Abstract

In complex maritime scenarios, ship infrared (IR) data shows diverse feature distributions due to the variation of target pose, behavior, and environmental factors. These factors pose challenges for feature learning in automatic target recognition (ATR) models. Furthermore, ATR models based on different mechanisms exhibit performance disparities, making it difficult to obtain a model that demonstrates optimal generalization across all scenarios. Therefore, this paper proposes a time-prior-based stacking ensemble deep learning model (TPSM) that integrates features extracted by multiple deep learning-based base models and leverages a meta model to inherit these features for complementary advantages. Simultaneously, to mitigate the adverse effects of complex spatial distribution within the dataset, the training set is partitioned into samples characterized by distinct time attributes (e.g., day, night, and mixed). Base models are then trained and selected from these diverse time attribute samples to construct the ensemble framework. The results indicate that TPSM achieved an Accuracy performance of 95.46% on high-fidelity simulated data, surpassing individual base models and other ensemble learning methods. Finally, we employ some explainable artificial intelligence (XAI) techniques to visualize the learning features of the TPSM, thereby validating the credibility of feature extraction.