基于深度神经网络的台风中心定位方法

A Typhoon Center Location Method Based on Deep Neural Network

关键词:  台风中心定位  注意力机制  神经网络  距离损失函数

Key words:  typhoon center location  attention mechanism  neural network  distance loss function

摘要:

台风中心定位的微小误差会对台风路径预报造成较大的偏离，因此精确定位台风中心是台风路径预测和灾害预报的重要步骤。台风云系随时间不断变化且风力强弱不一，在卫星云图中呈现了多样性和复杂性，现有基于神经网络的模型由于缺少对台风特征图像多维度参数的权重合理分配，在自动提取台风图像特征上受到了限制。为此，提出一种融合通道注意力与坐标注意力的神经网络模型（TY-LOCNet），首先搭建深度卷积神经网络模型提取台风特征；其次引入通道注意力机制从台风特征中捕获通道级别的信息，提升模型对重要通道的关注度；然后将通道注意力结果输入到坐标注意力机制中全局标定台风位置信息，使模型能够在较大的区域关注到台风的形态结构；此外，均方误差损失函数未能融合计算坐标导致定位精度低，因此提出距离损失函数(DISTLoss)通过距离回归提高模型定位精度。实验结果表明，TY-LOCNet的平均位置误差(MLE)、平均定位误差(MAE)和检测速度分别为 3.502 像素，0.292 °和 17 FPS，优于其他模型。台风中心定位模型 TY-LOCNet可为台风预报提供实时性台风中心定位支持。

Abstract:

Minor errors in typhoon center position can cause significant deviations in typhoon path prediction, so accurately locating typhoon center is an important step in typhoon path prediction and disaster early warning. Typhoon cloud systems change continuously with varying wind strength, leading to diverse and complex satellite images. Existing models based on neural networks are limited in the automatic extraction of typhoon features due to the lack of reasonable weight allocation for multi- dimensional parameters in typhoon images. For this reason, this paper proposed a neural network model (TY-LOCNet) that integrated channel attention and coordinate attention. Firstly, a deep convolutional neural network model was built to extract typhoon characteristics. Secondly, the channel attention mechanism was introduced to capture channel-level information from typhoon characteristics and enhance the attention of the model on important channels. Moreover, the channel attention results were input into the coordinate attention mechanism to calibrate typhoon position information globally so that the model can focus on the morphological structure of typhoons in large areas. Furthermore, the mean square error loss function failed to fuse the calculated coordinates, resulting in low locating accuracy. Therefore, the distance loss function (DISTLoss) was proposed to improve the locating accuracy of the model through distance regression. Experimental results show that the mean location error, mean absolute error, and detection speed of TY-LOCNet were 3.502 pixels, 0.292°, and 17 FPS, respectively, outperforming other models. Therefore, the typhoon center location model TY-LOCNet may provide real-time information on typhoon center position for typhoon forecasting.