loading page

Snowfall Type Classification for Improving Passive Microwave Snowfall Estimates
  • +1
  • Lisa Milani,
  • Veljko Petkovic,
  • Marko Orescanin,
  • Mark Kulie
Lisa Milani
Earth System Science Interdisciplinary Center (ESSIC) - University of Maryland

Corresponding Author:[email protected]

Author Profile
Veljko Petkovic
Earth System Science Interdisciplinary Center (ESSIC) - University of Maryland
Author Profile
Marko Orescanin
Navy Postgraduate School
Author Profile
Mark Kulie
Author Profile


Quantitative Precipitation Estimates from space-based observations represent an important dataset for understanding the Earth’s atmospheric, hydrological and energy cycles. Precipitation retrieval algorithms have been developed and refined over the last few decades and their accuracy and reliability are becoming increasingly more important for Earth’s energy budget and human activities. In particular, snowfall represents a key component of water cycle and contributes significantly to the Earth’s radiative balance. Accurately quantifying global surface snowfall is especially important since snow comprises a large percentage of the annual surface precipitation in many regions. The complexity of snowflakes particles and the nonlinear relationships between observations and retrieved variables have moved scientists to continually develop and enhance retrieval techniques. The present work aims to improve snowfall retrievals from Passive Microwave (PMW) sensors. Within the Global Precipitation Measurement (GPM) mission, the Goddard PROFiling (GPROF) algorithm snowfall retrieval has been chosen as an example of PMW precipitation product. Since previous works have demonstrated that GPROF performance strongly depends on the snowfall type, we developed a Machine Learning technique to classify snowing regime. A combined CloudSat-GPM dataset has been used to build the training dataset in which the GPM Microwave Imager (GMI) brightness temperatures (TB) are associated with a snowfall type, classifying the snowfall into three classes (‘shallow convective’, ‘deep stratiform’, ‘other’). The snowfall classification was adopted from a CloudSat classifying technique, based on snowing profiles and cloud classification. The problem is posed as a supervised learning problem, using a fully connected deep learning architecture with TB textures serving as input features. After building, optimizing and applying the classification method to existing GMI data, an evaluation exercise was undertaken to assess the classification performance. Results show that using only 20,000 GMI-CloudSat collocated snowing scenes the accuracy in retrieving the three classes has exceeded 80%. Being able to classify snowfall mode will help develop a specific setup for GPROF to improve detection and retrieval performance.