Fast Machine Learning-Based Extraction Of The Peak Number From Icesat Full-Waveform Data

As an important waveform feature, the number of peaks indicates the complexity of the surface coverage within the laser footprint. At present, the common methods of determining the number of peaks are based on waveform decomposition because the full waveform can be regarded as the superposition of multiple Gaussian waves. However, the empirical thresholds can affect the performance of the algorithms, and the process is time-consuming and inefficient. This study attempts to develop an automatic peak-number-driven classifier and picker based on machine learning (ML) and convolutional neural network (CNN) models. Waveforms were divided into four categories according to the peak-number. The results of five classifiers show good agreement with the verification data from GLAH14, with overall accuracies of 90.65%, 92.69%, 92.45%, 91.95%, and 93.05%. Moreover, F (1) score was calculated to evaluate the classification performance of each classifier for different categories. The results indicate that each classifier accurately distinguishes waveforms with different numbers of crests, especially invalid and single-peak waveforms. The ability to quickly and automatically extract the number of peaks based on ML methods will be of great benefit for specific research to screen suitable experimental data