基于多尺度纹理特征的海底底质样本增强方法

Sample enhancement methods based on multiscale texture features applied to seabed sediment classification

针对多波束海底底质分类模型构建受限于样本和特征对海底底质类型表征不足、模型稳定性差等问题,利用多尺度滑动窗口法提取声学纹理特征,结合K-均值聚类分析其精度,实现了多尺度纹理特征的优选,并利用多尺度纹理特征并辅以地形特征结合SLIC样本增强方法,实现了底质样本的有效扩充。同时,利用随机森林、BP神经网络、K最邻近、支持向量机等4种经典监督分类模型训练预测和评估所扩充的样本数据,最终总体分类精度均达到90%以上,kappa系数达到0.85以上。

To address the problems that the construction of multibeam seabed sediment classification model is limited by the insufficient characterization of seabed sediment types by samples and features and the poor stability of the model,this paper uses the multi-scale sliding window method to extract acoustic texture features.The preference of multi-scale texture features are also achieved by combining the K-means clustering to analyze its accuracy.The sediment samples are expanded effectively using multi-scale texture features and supplements them with topographic features combined with the SLIC sample enhancement method.Meanwhile,four classical supervised classification models,including Random Forest,Back Propagation Neural Network,K-Nearest Neighbor,and Support Vector Machine are trained to predict and evaluate the expanded sample data.The result show that final overall classification accuracy reaches more than 90%,and the kappa coefficient reaches more than 0.85.