Student Research Reports
Open social media sharing options menu
Mesoscale Precipitation Nowcasting experiment based on the ConvLSTM Model (neural networks)
Organization(s):The Affiliated High School of National Chung Hsing University
Country:Taiwan Partnership
Student(s):Chi Hsiao, Yi-Ta Tsai, Ting-Yi Lai, and Ying-Lung Liu
Grade Level:Secondary School (grades 9-12, ages 14-18)
GLOBE Teacher:Shih-chao Lin
Contributors:Cxxixi, Ministry of Science and Technology & Chinese Culture University, Data Bank for Atmospheric and Hydrologic Research
Report Type(s):International Virtual Science Symposium Report
Protocols:Precipitation
Presentation Poster:
View Document
Optional Badges:I am a Collaborator, I am a Data Scientist, I work with a STEM Professional
Language(s):English
Date Submitted:03/10/2020
Climate change has increased the rate of short-term severe weather change and the difficulty and uncertainty of weather nowcasting. Therefore, the way to predict short-term rainfall with precision has become an important research topic.
This research is aimed to increase the precision of precipitation nowcasting by combining image processing with neural networks in machine learning. The ConvLSTM model we used is trained with large amounts of doppler radar images, and by using the Marshall–Palmer relation, we can calculate the converted rainfall at a given point, thus predicting the rainfall.
Results show that the model has over 60% hit rate up to forecasting 60 minutes to the future when trained with 30 days of radar images for 250 epochs. And through calculating the values of a and b in the Marshall–Palmer relation (dBZ=10×log(a)+10(b)×logR) below the altitude of 800 meters in central Taiwan, we found that with a equaled to 87.97±0.35 (R^2=0.52) and b equaled to 1.32, the model can predict rainfall the most accurately.