Abstract: Deep learning-based object detection in low-altitude scenarios is susceptible to environmental disturbances and decision-making opacity, which introduces unacceptable unknown risks to airborne systems. To address this, this paper proposes a methodology for the explainability analysis and explanation quality evaluation of deep learning models in low-altitude object detection. First, data resources related to low-altitude object detection are systematically reviewed to construct a fused dataset encompassing diverse backgrounds and multiple object categories; subsequently, mainstream deep learning object detection algorithms, including YOLO, RCNN, and DETR, are employed to train low-altitude object detection models. Second, the SHAP algorithm is introduced to calculate the marginal contributions of input features and their interaction terms to the model's predictive outcomes, enabling precise feature attribution and visual explanation of the critical discriminative regions within the samples. Finally, an explanation quality evaluation metric based on a likelihood ratio is proposed to establish a quantifiable, comparable, and verifiable measurement criterion for the models, and the detection results are calibrated utilizing explainability prior information. Experimental results demonstrate that the DETR model achieves optimal performance in low-altitude object detection, with precision and mAP@50 reaching 90.54% and 83.56%, respectively. The explainability analysis further reveals a positive correlation between model performance and explanation quality, and the calibration of low-altitude object detection driven by explainability priors significantly enhances the detection performance of highly capable models, thereby validating the feasibility and effectiveness of the proposed method.

Key words: low-altitude economy, deep learning, explainability, intelligent system safety, airworthiness