Target Visibility Prediction Method Based on Cost-Sensitive CNN-BiLSTM Network

To address the low accuracy and insufficient autonomy of the traditional SGP4 model in predicting target visibility under non-perfect ephemeris conditions, a cost-sensitive hybrid network-based method was proposed for predicting the visibility of high-dynamic targets in specific regions. This method employed an early fusion of multi-source heterogeneous data features and constructed a CNN-BiLSTM architecture to comprehensively extract spatial geometric local features and temporal short-term dependency features. A cost-sensitive learning mechanism was introduced to address the extreme imbalance in target visibility, enabling accurate prediction of visibility windows in specific regions. The results showed that with an extreme imbalance ratio of 1:357, the proposed method achieved a precision of 94.56% and a recall of 97.78%, outperforming existing methods. This approach provides technical support for space mission planning and collision avoidance warning.