A COMPARISON OF CONVOLUTIONAL NETWORKS CAPSULE NETWORK AND TRANSFER LEARNING FOR HEAVY A INFALL NOWCASTING


Ozden C.

FRESENIUS ENVIRONMENTAL BULLETIN, vol.31, no.1A, pp.1103-1110, 2022 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 31 Issue: 1A
  • Publication Date: 2022
  • Journal Name: FRESENIUS ENVIRONMENTAL BULLETIN
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Agricultural & Environmental Science Database, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), CAB Abstracts, Chemical Abstracts Core, Communication Abstracts, Environment Index, Geobase, Greenfile, Metadex, Pollution Abstracts, Veterinary Science Database, Civil Engineering Abstracts
  • Page Numbers: pp.1103-1110
  • Keywords: Nowcasting, deep learning, capsule network, transfer learning
  • Çukurova University Affiliated: No

Abstract

This paper studies the efficiency of different deep learning architectures in nowcasting heavy rainfalls. For this purpose, heavy rainfall records were collected from 2014 to 2020 for Antalya, Turkey. Radar images and actual weather charts were collected from the archives of Turkish State Meteorological Services. Four different models i.e., CNN, ConvGRU, CapsNet and EfficientNetB07 were trained and tested on the curated dataset. The results established the presence of a pattern in the distribution of heavy rainfalls in the study area. CapsNet and ConvGRU models attained comparatively lower accuracies, EfficientNetB07 pretrained model performed best with over 84% accuracy in detecting the heavy rainfall incident 3 hours earlier.