Research on LiDAR Clear Air Turbulence Recognition Based on Improved SE-ResNet50

To address the issue of LiDAR’s low turbulence recognition rate at airports in low-altitude areas, a clear air turbulence recognition method based on an improved Squeeze-and-Excitation Residual Network with 50 layers (SE-ResNet50) is proposed. By introducing the squeeze-and-excitation module and improving the network structure, the model’s excessive sensitivity to feature location is reduced, thereby enabling the network to selectively highlight useful information features during the learning process. A sample dataset was established using measured data from Lanzhou Zhongchuan International Airport; for model training, a balanced dataset was created by extracting an equal amount of weak, moderate, and strong turbulence data based on the turbulence classification level. Under the same experimental conditions, the recognition accuracy of the improved SE-ResNet50 was increased by 7.44%, 6.52%, and 4.11% compared with the convolutional neural network, MobileNetV2, and ShuffleNetV1 networks, respectively. A comparison of the confusion matrices generated by each model showed that the accuracy of the proposed method reached 95%, verifying the feasibility of the proposed method.