DOA Estimation of Wideband LFM Sources based on Narrowband Methods Integration Using Random Forest Regression

For underwater target detection, DOA estimation of wideband LFM sources is one of the most important and challenging issues, which should be solved efficiently and reliably. However, in practice, especially under low Signal-to-Noise-Ratio (SNR) condition, DOA estimation performance of the mainstream methods, such as Incoherent Signal Subspace Method (ISSM), Coherent Signal Subspace Method (CSSM), and Fractional Fourier Transform (FrFT) based Method, is always unsatisfactory in both accuracy and false alarm rate. In this paper, at first, a group of spatial spectrums were obtained by employing several complementary narrowband DOA estimation methods like the one for each sub-band of frequency snapshots in the ISSM method respectively. And then, using random forest regression (RFR) technique, the obtained spatial spectrums were encoded to the expected one sequentially from lateral to longitudinal directions. Finally, DOA information can be determined using the obtained two-stage encoded spatial spectrum. Experimental results illustrated that there is indeed complementarity among the employed narrow-band DOA estimation methods, and this complementarity can be exploited to significantly improve the performance of the traditional ISSM method.