Cloud Detection Methods For Millimeter-wavelength Radar Based On Computer Vision

Clouds play an important role in weather and climate change,and are one of the biggest uncertainties in climate projection.Accurate long-term observation of clouds is essential to validate and constrain the climate models,and reduce the uncertainty of climate projection.Millimeter-wavelength cloud radar can obtain the three-dimensional structure of cloud,by penetrating cloud layer and receiving the echo signal backscattered from cloud droplet.It is a powerful tool to detect clouds,and thus has been widely used in ground-based and spaceborne platforms,to obtain the detailed information about the vertical structure,radiative effect,and climate impact of clouds.In practice,the raw radar data first need to be processed to identify the cloud signals,i.e.,cloud detection.In addition,considering the variance in radiative effects and climate impact of different types of clouds,the specific types of clouds,e.g.,low-level liquid clouds,high-altitude ice clouds and deep convective systems,need to be identified for particular application scenarios.The raw radar received power can be somewhat regarded as a two-dimensional image,and there is a subdivision field in machine learning,that deals with how computers can gain high-level understanding from digital images or videos,in a way close to human beings,namely computer vision.Computer vision seeks to understand and automate tasks that the human visual system can do,by combining image processing,pattern recognition and artificial intelligence technology.It focuses on the calculation and analysis of one or a series of images to realize the recognition of the target object,determine the position and attitude of the target object,and make symbolic description and interpretation of the threedimensional scene.Its principle is similar to the remote sensing image interpretation.Therefore,based on the relevant algorithms in computer vision,this thesis uses the ground-based and spaceborne radar observations,to improve and apply the relevant algorithms for the practical issues of millimeter-wavelength radar cloud detection,such as weak signal recognition,low-cloud and clutter separation,and deep convective system identification.The main contents and conclusions of this thesis are as follows:(1)The traditional cloud detection algorithm is improved by using the basic idea of bilateral filter from computer vision.Through the pre-judgment of cloud signal and noise,the strong radar received power and noise are firstly separated.Then,considering the high-frequency characteristics of noise and the low-frequency spatial continuity of cloud signal,a bilateral filter is constructed to process the them respectively,which effectively compresses the noise level,reduces the overlapping area of noise and cloud signal,and identifies more weak echo signals.Both theoretical analysis and simulation results show that the algorithm not only compresses the radar background noise,but also maintains the clear edge between the weak signal and noise,and accurately identifies more real signals ignored by previous algorithms.Aiming at the Ka-band zenith radar(KAZR)at Semi-Arid Climate and Environment Observatory of Lanzhou University(SACOL),a cloud detection algorithm based on bilateral filtering noise compression is proposed,and it is proved that the cloud detection results have obvious improvement on weak signal areas such as thin cirrus and ice cloud top.(2)A low-cloud and clutter separation algorithm for KAZR is proposed by using multi-dimensional probability distribution functions along with the Bayesian method.The radar reflectivity,linear depolarization ratio,spectral width,and their dependence on the time of the day,height,and season are used as the Bayesian discriminants.A low-pass spatial filter is applied to the Bayesian undecided classification mask by considering the spatial correlation difference between clouds and clutter.Analysis of typical individual cases shows that the algorithm can separate most of the clutter and still retain the low cloud signal,and demonstrates good discrimination even when the low-cloud signal is mixed with clutter.The identification results of the algorithm are evaluated by using the observation from millimeter-wavelength radar and lidar for one year.It is found that the detection rate is 98.5 % and the false alarm rate is 0.4 %,which meets the actual needs.(3)The cloud detection algorithm of the spaceborne millimeter-wave cloud radar,i.e.,the cloud profiling radar carried on Cloud Sat,was improved using the basic idea of bilateral filter from computer vision.Combined with official algorithm's own alongtrack integration scheme,the signal and noise are processed separately using the bilateral filter,while compressing the background noise.The results of the improved algorithm have less background noise than the official R04 version,that is,less false detection rate,and more real signals are detected than the official R05 version,that is,less false omission rate.Overall,the method balances the false detection rate and the false omission rate,and ultimately improves the accuracy of cloud detection products,especially in the intertropical convergence zone and mid-high latitude storm track zones.(4)The connected component analysis method in computer vision is used to identify the cloud objects from the Cloud Sat observation image,and the deep convective system is identified according to the cloud type products.The identified deep convective systems are then divided into the convection pillar,the major anvil and the minor anvil.The cloud top buoyancy is calculated to determine the life stages of deep convective systems,namely,the developing,mature and dissipating stages.The morphology and microphysical structure of the deep convective system during different life stages are demonstrated by synthetic analysis.(5)Taking the radar track detection as an example,the potential application scenarios of bilateral filter in computer vision to other remote sensing fields are preliminarily discussed,combined with Hough transform.From the simulation result,it is initially confirmed that the bilateral filter method has good portability and universality.