基于改进DeeplabV3+的HFSWR电离层杂波及海杂波自动识别

Automatic recognition method of ionospheric clutter and sea clutter for High Frequency Surface Wave Radar based on improved DeeplabV3+

摘要:

针对高频地波雷达(HFSWR)回波谱中背景噪声复杂、杂波占比较小且电离层杂波形态位置各异，难以自动识别的问题，以DeeplabV3+深度学习算法为基础架构并加以改进，提出一种HFSWR电离层杂波及海杂波自动识别方法。选用轻量级MobileNetV2作为主干特征网络，加入通道注意力机制模块SENet，实现对杂波标签的侧重学习，优化训练集中各类标签的损失权重；采用模型预训练迁移法对网络进行预训练，解决样本空间过小的问题。实测数据集上的实验结果表明，本方法可以实现HFSWR电离层杂波及海杂波的自动识别。相比原DeeplabV3+算法，杂波识别结果更为准确和精细，海杂波识别结果的平均交并比(mIoU)和准确率(ACC)分别提升2.9%和5.1%；电离层杂波识别结果的mIoU和ACC分别提升3.0%和4.9%。

Abstract:

The background noise in the echo spectrum of High Frequency Surface Wave Radar(HFSWR) is complex, the clutter accounts for a small proportion and the ionospheric clutter has different forms and positions, therefore, it is difficult to automatically recognize the clutters. Based on DeeplabV3+ deep learning algorithm, an automatic identification method of ionospheric clutter and sea clutter is proposed for HFSWR. Selecting the lightweight MobileNetV2 backbone feature network, adding the channel attention mechanism module SENet, the focused learning of clutter labels is realized, and the loss weight of various labels in the training set is optimized. The model pre-training transfer method is employed to pre-train the network to tackle with the problem of too small sample space. The experimental results on the measured data set show that the proposed method can realize the automatic recognition of ionospheric clutter and sea clutter in HFSWR, and can obtain more accurate and finer clutter recognition results than the original DeeplabV3+ algorithm. The mean Intersection over Union(mIoU) and Accuracy(ACC) of sea clutter recognition results are increased by 2.9% and 5.1% respectively, and the mIoU and ACC of ionospheric clutter recognition results are increased by 3.0% and 4.9% respectively.