College of Geomatics, Shandong University of Science and Technology, Qingdao 266590, China.
Key Laboratory of Submarine Geosciences, Second Institute of Oceanography, Hangzhou 310012, China.
Sensors (Basel). 2018 Nov 8;18(11):3828. doi: 10.3390/s18113828.
Airborne light detection and ranging (LiDAR) full waveforms and multibeam echo sounding (MBES) backscatter data contain rich information about seafloor features and are important data sources representing seafloor topography and geomorphology. Currently, to classify seafloor types using MBES, curve features are extracted from backscatter angle responses or grayscale, and texture features are extracted from backscatter images based on gray level co-occurrence matrix (GLCM). To classify seafloor types using LiDAR, waveform features are extracted from bottom returns. This paper comprehensively considers the features of both LiDAR waveforms and MBES backscatter images that include the eight feature factors of the LiDAR full waveforms (amplitude, peak location, full width half maximum (FWHM), skewness, kurtosis, area, distance, and cross-section) and the eight feature factors of MBES backscatter images (mean, standard deviation (STD), entropy, homogeneity, contrast, angular second moment (ASM), correlation, and dissimilarity). Based on a support vector machine (SVM) algorithm with different kernel functions and penalty factors, a new seafloor classification method that merges multiple features is proposed for a beneficial exploration of acousto-optic fusion. The experimental results of the seafloor classification around Yuanzhi Island in the South China Sea indicate that, when LiDAR waveform features are merged (using an Optech Aquarius system) with MBES backscatter image features (using a Sonic 2024) to classify three types of sands, reefs, and rocks, the overall accuracy is improved to 96.71%, and the kappa reaches 0.94. After merging multiple features, the classification accuracies of the SVM, genetic algorithm SVM (GA-SVM) and particle swarm optimization SVM (PSO-SVM) increase by an average of 9.06%, 3.60%, and 2.75%, respectively.
机载激光探测和测距（LiDAR）全波形和多波束回声探测（MBES）反向散射数据包含有关海底特征的丰富信息，是代表海底地形和地貌的重要数据源。目前，在使用 MBES 对海底类型进行分类时，从反向散射角响应或灰度中提取曲线特征，并基于灰度共生矩阵（GLCM）从反向散射图像中提取纹理特征。在使用 LiDAR 对海底类型进行分类时，从底部回波中提取波形特征。本文综合考虑了 LiDAR 波形和 MBES 反向散射图像的特征，包括 LiDAR 全波形的八个特征因素（幅度、峰值位置、半峰全宽（FWHM）、偏度、峰度、面积、距离和横截面）和 MBES 反向散射图像的八个特征因素（均值、标准差（STD）、熵、同质性、对比度、角二阶矩（ASM）、相关性和相异性）。基于具有不同核函数和惩罚因子的支持向量机（SVM）算法，提出了一种融合多种特征的新的海底分类方法，以有益地探索声光融合。南海元志岛周围海底分类的实验结果表明，当将 LiDAR 波形特征（使用 Optech Aquarius 系统）与 MBES 反向散射图像特征（使用 Sonic 2024）融合以对三种类型的沙、珊瑚礁和岩石进行分类时，总体精度提高到 96.71%，kappa 达到 0.94。融合多种特征后，SVM、遗传算法 SVM（GA-SVM）和粒子群优化 SVM（PSO-SVM）的分类精度平均提高了 9.06%、3.60%和 2.75%。
