Researches On GPR Shallow Target Detection Based On Hierarchical Clustering Algorithm And Random Forest

The construction and maintenance of underground pipelines have promoted the development of shallow target detection technology of ground-penetrating radar(GPR).During the detection process,a lot of noise and clutter will generate,which it's amplitudes or waveforms are similar to the pipeline target.If the detection process is not robustness enough for the noise and clutter,these interferences will lead to several problems such as misperception or missing target while detect pipeline.Therefore,this paper aims to research a GPR shallow target detection method,which is more robust against noise and clutter,based on the accuracy of hierarchical clustering for obtaining data distribution,as well as the strong robustness and generalization ability of random forests.As for the problem that the GPR shallow target detection process is greatly affected by clutter,a multi-label hierarchical clustering based detection method(MHCD)is proposed.First,the multi-label hierarchical clustering method is constructed by combining information entropy and agglomerative hierarchical clustering method(AHC).Then,extracting target signals based on the shape and texture feature of the clustering result.Compared with the traditional unsupervised clustering method,MHCD uses the instability between data as a similarity measure.Besides,it can obtain the difference between the target data and the distribution of clutter data more accurately so that the detection process can be more robust against the clutter with similar amplitude but different waveforms to target signals.Finally,due to the calculation of MHCD only needs to take into account the similarity with the adjacent data,which can speed up the data processing.In order to further improve the robustness of the noise and the clutter with similar waveforms to target signals in the detection process,this paper applies random forest(RF)to segment GPR images based on its strong robustness and generalization ability,and a method by combining RF and MHCD(RFMHCD)is proposed for detecting GPR shallow targets.First,based on the correlation between the pixel level and the vertical direction of the neighborhood,the crisscross decussation feature extraction method(CCDF)is constructed for pixel classification.Then the ensemble learning model RF-SAE based on RF and sparse autoencoder(SAE)is constructed for pixel classification,and through the constrained K-means optimization model parameters to improve the classification performance.Finally,a pixel regionalization method based on the geometric features of the target signal is constructed,and is utilized to segment GPR images according to the pixel classification results to improve the completeness of the target.Compared with the traditional image segmentation method,the operation of segmenting GPR images based on RF-SAE improves the robustness against noise.Besides,the results segmented by combing RF-SAE and pixel regional method has a high degree of integrity,on this basis,combing MHCD and the random forest-based image segmentation method can further improve the robustness of the detection process to noise and waveform-like clutter.