Improved Forest Signal Detection for Space-borne Photon-counting LiDAR Using Automatic Machine Learning

NASA's (National Aeronautics and Space Administration) ICESat-2 with a Photon Counting LiDAR (Light Detection And Ranging) Sensor was successfully launched in September 2018. The sensor uses an advanced detection system called the Advanced Topographic Laser Altimeter System (ATLAS). The ATLAS sensor detects signal photons at high speed and is highly sensitive. However, the sensor also extracts a large amount of background photon noise coming from the atmosphere, ground, sun, or other radiation. This condition is particularly evident in forest areas. Therefore, the method of filtering noise is of great significance for any use of the data. Without human intervention, automatic machine learning can form a set of processes needed for classification, namely feature selection, model selection, and model evaluation. This method offers convenient calculation, transferability, applicability, and interpretability. This study proposes an automatic machine learning approach to utilize data for forestry applications to improve data availability compared to NASA's official product. We used only 10% of the sample points for training on five datasets in the forest region and compared the performance of the classifiers. First, we conclude that the integrated learning performance generally outperforms single models, and the mean F1 score of all tests is approximately 0.9. The mean F1 score of the Stacked Ensembles model is 0.957 ahead of the other models. The top three variables used in training models are kNNdist5, kNNdist10, and h. These three variables could explain 51.6% of the components of the models. Over the regions tested, the proposed method could improve the proportion of signals correctly identified by 6.4%, 12.2%, 2.7%, 9.3%, and 1.4% in five datasets. The model performs better in low signal-to-noise (SNR) datasets less than 7.5. Then, compared to distinguishing noise photons, the optimal classifiers did better classifying signal photons from noise. The classifiers could correct misclassified labels in ATL08 products and show good stability in different conditions. A new method for the separation of forest signal from noise has been demonstrated, which uses only a very limited number of sample points for training, ensuring operational efficiency and training accuracy. The method would be largely unaffected by differences in topography, noise distribution, and SNR. Moreover, the classifiers demonstrated the ability to correctly identify signals considered noise