Detection of cloud phase using ceilometer observations in New Zealand and the Southern Ocean.
Supercooled liquid water (SLW) clouds are commonly observed in the Southern Hemisphere but are poorly represented in current climate models. Due to signal attenuation, satellite borne lidar observations of SLW are biased, so surface-based measurements are needed for understanding the distribution of...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10092/106162 https://doi.org/10.26021/15045 |
| Summary: | Supercooled liquid water (SLW) clouds are commonly observed in the Southern Hemisphere but are poorly represented in current climate models. Due to signal attenuation, satellite borne lidar observations of SLW are biased, so surface-based measurements are needed for understanding the distribution of SLW cloud for the purposes of model evaluation. The use of depolarization lidar to measure atmospheric volume depolarization ratio (VDR) is a common technique in classifying cloud phase (liquid or ice). Unfortunately, such lidars are uncommon and not suited to remote and extended measurement campaigns. This thesis is focused on identifying whether ceilometers (simple, low-power lidars) can be utilized to quantify SLW occurrence over the Southern Hemisphere, particularly the Southern Ocean. Previous work has established a technique to detect supercooled liquid water containing clouds (SLCC) from ceilometer retrievals, using a supervised machine learning model. Utilizing a depolarization lidar training dataset from Davis Station, Antarctica, that study trained an eXtreme Gradient Boosting (XGBoost) model on a set of copolarized attenuated backscatter peak properties. However, no warm liquid water clouds (WLCC) were present in the Antarctic training dataset, limiting the transferability of that model (hereafter named G22-Davis) to other regions where WLCC is present. In this study, we apply and evaluate G22-Davis on a 9-month Micro Pulse Lidar (MPL) dataset collected in Christchurch, New Zealand, a mid-latitude site. After building a reference VDR cloud phase mask, we found that G22-Davis performed relatively poorly at SLCC detection with an accuracy of 0.62. Unsurprisingly, G22-Davis often misclassified WLCC as SLCC. We then trained a new model, G22-Christchurch, to perform SLCC detection on the same set of co-polarized attenuated backscatter peak properties. G22-Christchurch performed well, with accuracy scores as high as 0.89. To interpret the model results, we analysed feature importance scores and found that using ... |
|---|