U-NET CNN ARCHITECTURE AND OCR FOR HIGH-ACCURACY PATTERN AND TEXT RECOGNITION IN METEOROLOGICAL FACSIMILE DATA

Abstract

The categorization of meteorological data through imagery plays a vital role in weather prediction, environmental surveillance, military activities, target identification, high-altitude aviation, image retrieval, and maritime navigation. The study explores the application of machine learning through convolutional neural networks (CNNs) to derive crucial and relevant insights from weather facsimile charts. The data on weather facsimile charts is frequently identified and depicted on weather charts, relying on the expertise of forecasters. Identifying unique weather patterns from linear weather systems presents a significant challenge in the field of meteorological science research. This paper presents a significant contribution to pattern recognition and meteorological computing through the introduction of the U-Net Convolutional Neural Network (CNN) classification model, showcasing the effectiveness of this methodology in analyzing meteorological information. Utilizing a framework for hyper-parameter optimization based on convolutional layers, our CNN system identifies weather patterns with an accuracy ranging from 89% to 99% (Dashed Line pattern, Weather symbols). Recently, based on character recognition, Optical Character Recognition is commonly utilized in many straightforward applications. This paper presents Easy OCR, a deep learning library based on Python that includes detection models for over 85 languages. An accessible OCR method is proposed for character recognition from the weather facsimile maps dataset. The experiment demonstrates that the proposed method achieves an impressive accuracy rate of 95%-98% in the thorough and precise extraction of both textual and numerical data from weather facsimile charts.

Keywords: Meteorological facsimile charts; CNN; Data augmentation; Easy OCR, Pattern Recognition

https://doi.org/10.5281/zenodo.17818646