In the remote detection of bioaerosol clouds by fluorescence lidar, the decision tree method is often used to identify the fluorescence spectral signals of the clouds. The conventional decision tree algorithm selects the intensity values of the echo signals at different wavebands as features rather than extracting the statistical features of the echo signals, thereby effectively recognizing the fluorescence spectra measured under the same environmental conditions. However, in bioaerosol LiDAR, the acquired fluorescence spectra are highly variable because of the great uncertainty of the atmospheric state and background radiation, such that when the signal-to-noise ratio of LiDAR decreases, the previously established decision tree model may be overfitted, resulting in low recognition accuracy. In this study, the conventional algorithm is improved to increase the noise resistance of recognition and make the algorithm applicable to the field of LiDAR detection of bioaerosols.

In this study, fluorescence spectral signals of six biomaterials are first tested under laboratory conditions. Different Gaussian white noises with different intensity values are added to the fluorescence spectrum of each material to simulate the actual echo signals detected by bioaerosol LiDAR. Subsequently, the fluorescence spectra and recognition algorithms are analyzed mechanistically, and a decision tree recognition algorithm based on statistical feature extraction is designed, primarily based on discrete cosine transform (DCT), central peak position, and normalized spectral area. Finally, the performance of the two recognition algorithms is examined with simulated LIDAR signals under different noise intensity values. The two algorithms are used to train the spectra of the training set to form their respective decision trees, concurrently recording the training time. The decision trees are used to discriminate the test set, whereby the accuracy is calculated to analyze the actual detection ability of the algorithms before and after the improvement.

Both algorithms accurately recognize each biomass when the signal-to-noise ratio (SNR) of the signal is high. The recognition rate is above 90% when the SNR is above 20. However, the performance of the traditional algorithm dramatically weakens with an increase in noise. In the detection of bioaerosol LiDAR, the SNR is 10, leading to greatly reduced recognition accuracies of the traditional algorithms. The recognition accuracy of rapeseed pollen is lower than 60%. When the SNR is 5, the recognition accuracies are even lower than 50% for the four kinds of substances, clearly making it difficult to support the performance of the algorithms to meet the requirements of LiDAR telemetry. The improved algorithm maintains a recognition accuracy of above 65% even when the SNR is 5, and the recognition accuracy is above 80% when SNR is 10. Second, the training time of the algorithm designed in this study is 16?32 ms, which is much smaller than that of the traditional algorithm. This training time does not increase with the noise intensity, whereas the training time of the traditional algorithm, which is 84?509 ms, sharply increases with the noise intensity.

To solve the problem of efficient recognition of biofluorescence spectra by bioaerosol LiDAR, this study designs a novel decision tree algorithm based on statistical feature extraction of fluorescence spectra, by transforming the original primary multiple features into seven main high-level features through DCT, searching for the central wavelength, and calculating the spectral area, which covers almost all the spectral information. The proposed algorithm is faster to train and more noise-resistant, outperforming the traditional algorithm in all aspects. The results show that the decision tree algorithm based on feature extraction improves recognition accuracy and training speed, thereby averting misclassification and enhancing the detection performance of bioaerosol LiDAR.