Atmospheric Measurement Techniques (May 2024)
Deep-Pathfinder: a boundary layer height detection algorithm based on image segmentation
Abstract
A novel atmospheric layer detection approach has been developed based on deep learning techniques for image segmentation. Our proof of concept estimated the layering in the atmosphere, distinguishing between pollution-rich layers closer to the surface and cleaner layers aloft. Knowledge of the spatio-temporal development of atmospheric layers, such as the mixing boundary layer height (MBLH), is important for the dispersion of air pollutants and greenhouse gases, as well as for assessing the performance of numerical weather prediction systems. Existing lidar-based layer detection algorithms typically do not use the full resolution of the available data, require manual feature engineering, often do not enforce temporal consistency of the layers, and lack the ability to be applied in near-real time. To address these limitations, our Deep-Pathfinder algorithm represents the MBLH profile as a mask and directly predicts it from an image with backscatter lidar observations. Deep-Pathfinder was applied to range-corrected signal data from Lufft CHM15k ceilometers at five locations of the operational ceilometer network in the Netherlands. Input samples of 224 × 224 px were extracted, each covering a 45 min observation period. A customised U-Net architecture was developed with a nighttime indicator and MobileNetV2 encoder for fast inference times. The model was pre-trained on 19.4×106 samples of unlabelled data and fine-tuned using 50 d of high-resolution annotations. Qualitative and quantitative results showed competitive performance compared to two reference methods: the Lufft and STRATfinder algorithms, applied to the same dataset. Deep-Pathfinder enhances temporal consistency and provides near-real-time estimates at full spatial and temporal resolution. These properties make our approach valuable for application in operational networks, as near-real-time and high-resolution MBLH detection better meets the requirements of users, such as in aviation, weather forecasting, and air quality monitoring.