Identification of military and civilian targets by LSTM

This study presents an artificial intelligence-based approach for classifying military and civilian air targets as friendly, hostile, or unknown. Conventional methods widely used in operational environments rely on the rule-based IDCP (Identification Data Combining Process) defined under NATO STANAG 4162, which utilizes expert-driven deterministic decision mechanisms through structures such as the Source Probability Matrix (SPM), Mapping Matrix (MM), and Force Mix Ratio (FMR). Although these systems provide strong explainability and auditability, they show limitations in flexibility and scalability when sensor sources produce time-varying, missing, or contradictory information. The proposed approach integrates multiple sources—radar detections, ESM signals, IFF responses, flight plans, MPR information, platform performance constraints (PP), and behavioral indicators (TAB)—into a unified time-series learning pipeline using an LSTM architecture. The preprocessing stage includes time synchronization, missing-data imputation, normalization, and scenario-based labeling to obtain a consistent model input representation despite varying sampling rates and reliability levels across sensors. The training process is performed entirely on simulated multi-source scenarios, and techniques such as noise injection, time shifting, and partial information masking are employed to enhance robustness under uncertain conditions. The LSTM architecture exhibits improved flexibility compared to the rule-based system in capturing temporal behavior changes, tracking anomalies such as sudden maneuvers or inconsistent IFF/ESM responses, and fusing heterogeneous information streams. Model performance is evaluated using standard classification metrics—accuracy, precision-recall, F1-score, and the confusion matrix—and shows more consistent results than the rule-based approach in scenarios involving temporal variations. However, the transparency and auditability advantages of IDCP remain valuable in specific cases. Overall, this study introduces a framework that combines the strengths of the STANAG rule-based structure with the adaptability of data-driven learning for time-dependent multi-source fusion. Future work will focus on expanding scenario sets for operational realism, deepening hybrid rule-learning integration, and reducing computational cost through model-lightening techniques.