The atmospheric boundary layer height (ABLH) is a critical factor in determining air pollution levels. Continuous observation of ABLH throughout the day and night is crucial for evaluating air quality. Light detection and ranging (LiDAR), one of the most traditional methods for measuring ABLH, offers high measurement accuracy and can provide detailed vertical atmospheric profiles with high temporal and spatial resolutions. However, there are several limitations. Current methods have low time resolution, with most sites operating only at specified times. Some sites provide data up to four times a day, which does not allow for continuous monitoring of the ABLH. LiDAR is an effective remote sensing method for detecting boundary layer height because it offers continuous measurements and provides high-resolution vertical atmospheric profile data. However, current LiDAR algorithms are prone to interference from complex atmospheric structures, such as cloud layers and suspended aerosols, which can affect the accuracy of boundary layer height detection under various conditions. Therefore, addressing these interferences is crucial. We propose an improved deep forest algorithm that integrates different remote sensing data to address the challenges associated with detecting boundary layer height using LiDAR.

We propose an improved deep forest algorithm that fuses multiple sources of remote sensing data. The optical data from micro-pulse LiDAR and the Doppler LiDAR, along with temperature, humidity, wind speed, and air pressure data from ground meteorological stations, are used to construct the dataset. We improve the deep forest algorithm in two main ways: 1) by employing feature selection methods to replace multi-dimensional scans, effectively removing redundant variables, and enhancing the dataset’s ability to capture relevant features; 2) a level-linked forest approach is used, where the input of each level in the linked forest is a combination of the output vectors from all previous levels and the original input feature vectors. The influence of the final result of the linked forest is evaluated, and the final output of the linked forest is used to replace that of the averaging method with a weighted approach. This adjustment aims to reduce the influence of weaker learning models on the outcome. Ultimately, by enhancing deep forest algorithms and applying them to train and predict the fused dataset, we obtain the final boundary height with improved accuracy.

The correlation coefficient between the boundary layer height obtained from the proposed method and the radiosonde measurements at the SGP site in 2020 is as high as 0.935 (Fig. 4). This is significantly higher than those of traditional methods like the gradient method and the threshold method, and it compares favorably with other machine learning algorithms. Case studies on clear days (Fig. 7) and cloudy days (Fig. 8) demonstrate that the results of the proposed method aligns closely with radiosonde measurements on clear days and remains unaffected by clouds and aerosols on cloudy days. To further validate the method’s performance, 103 cases of clear weather data (Fig. 7) and 50 cases of cloudy or aerosol weather data (Fig. 8) are analyzed. The results show that the proposed method improves the accuracy of ABLH retrieval on clear days and is robust against cloud and aerosol interference on cloudy days. In addition, an analysis of 377 daytime cases (Fig. 14) and 50 nighttime cases (Fig. 15) indicates that the method effectively improves the accuracy of ABLH retrieval during both day and night.

The improved deep forest algorithm, based on fused remote sensing data, significantly enhances the accuracy of LiDAR-based ABLH retrieval, achieving a correlation coefficient as high as 0.935 with radiosonde data. The method is effective in tracking the diurnal variation of the boundary layer height. Case analyses under different weather conditions demonstrate that the proposed method is robust and unaffected by complex atmospheric structures or nighttime conditions, providing reliable ABLH measurements.