Research On Weak Sea-surface Target Detection With Machine Learning-based Approaches

Sea-surface target detection has become an important way to obtain sea-surface information,which plays a decisive role in both military and civilian fields.There are several types of targets under the monitoring of the marine surface surveillance radars,such as the small-scale targets(small-sized boats),stealthy targets(stealth submarines),highmaneuvering targets(cruise missiles),and distant targets(remote sea rescue targets).As one of the common points,the targets above usually have low observable characteristics,collectively referred to as weak targets.The radar returns of these targets have low Signalto-Clutter radio(SCR)due to the low observability of the weak target and the complexity of the sea clutter.Therefore,an important question that arises is how to effectively detect the weak targets from the radar returned signals.In this paper,we construct a high distinguishable feature space,and design a false alarm rate included SVM-based detector when identifying the weak targets within sea clutter for achieve satisfactory classification results.For the construction of feature space,it is found that,compared with sea clutter,target signals show higher level of non-stationarity in the amplitude distribution.This difference can be quantified by the information entropy.Therefore,this paper adopts the information entropy as a distinguishable feature to establish a one-dimensional feature space.However,the extremely hostile marine environment will produce a large number of sea spikes in the sea clutter,which will significantly increase the non-stationarity of the sea clutter.This may further weaken the gap of the value of information entropy between the target signals and sea clutter.In view of this,we further combine the information entropy and Hurst exponent to construct a two-dimensional feature space.This directly follows the insight that the fractal shape of radar echo signals can avoid the influence of sea spikes on the target signals to some extent,which has been proved in existing literatures.Interestingly,through the discrete Fourier transform,we also found that the peaks of target signal are more convergent than that of the sea clutter in the frequency domain.More importantly,we devise an index to quantify this characteristic,referred to as the frequency-domain peak-to-average ratio(FPAR).At this point,the three-dimensional feature space is built by integrating the information entropy,Hurst exponent,and FPAR together.Then,we test the performance of the three-dimensional(3-D)feature space on the IPIX radar datasets.The obtained results show that the majority of the data are linearly separable by the feature space,while other data show nonlinear separability.These phenomena indicate that the detector is required to be capable of handling both linearly separable and linearly indivisible data.As a classical machine learning method,support vector machines(SVM)have the advantages of achieving satisfactory classification results for both linearly separable or linearly inseparable data.Although there have already been some works to construct SVMbased detectors to distinguish target signals from sea clutter,none of them have explored the detection probability of weak targets under quantitative false alarm probability.To address this issue,we in this paper design an algorithm that enables the SVM to work with a given false alarm probability,and study the detection performance of the SVM-based detector for an arbitrarily given false alarm probability.Moreover,the IPIX radar datasets are used to test and analyze the effects of different factors on the detector's performance,including the signal-to-clutter ratio,sampling duration,and different training data.The obtained results show that,compared with some existing classical algorithms,the detector designed in this paper significantly improves the detection probability of weak targets.