Deep feature extraction and explanation consistency analysis for fine-grained aircraft type classification in remote sensing imagery

Fine-grained aircraft type recognition in remote sensing imagery is of great value for military intelligence analysis and battlefield situation awareness. However, the limited transparency of deep learning decision-making processes restricts its application in high-stakes decision-making scenarios. To address this issue, this paper proposes an explainable aircraft type classification framework, termed EC-AT, which integrates transfer learning, explanation consistency analysis, and trustworthiness evaluation. Specifically, eight representative aircraft categories are selected from the MAR20 dataset to evaluate multiple deep neural network models under two settings: transfer-learning fine-tuning and random initialization. By leveraging pretrained model parameters and fine-tuning the classification layers, the proposed framework effectively improves feature representation capability and convergence efficiency under limited and class-imbalanced data conditions. To further interpret model decisions, three representative explainable artificial intelligence methods, namely Grad-CAM, LIME, and RISE, are introduced to analyze discriminative image regions from complementary perspectives, including class-discriminative activation, local perturbation-based approximation, and randomized input sampling. In addition, an IoMin-based spatial overlap metric is designed to quantitatively assess the consistency among regions generated by different explanation methods. Beyond explanation consistency analysis, the proposed framework is further evaluated using the Multisource AI Scorecard Table (MAST), which provides a structured assessment of model trustworthiness from multiple dimensions, including accuracy, consistency, logical argumentation, and visualization. Experimental results show that the transfer-learning-based models achieve excellent classification performance, with classification accuracy reaching up to 99.9%. The explainability analysis demonstrates that the proposed framework can identify key aircraft-related regions while revealing potential background interference and decision biases. Furthermore, the explanation consistency analysis and MAST-based evaluation jointly provide objective evidence for explanation results and the overall trustworthiness of the framework. Overall, the proposed EC-AT framework significantly improves the classification performance and explainability of fine-grained aircraft recognition in remote sensing imagery, offering a useful reference for trustworthy intelligent remote sensing analysis.