Real-Time Weather Classification using Transfer Learning for Enhanced Safety and Decision-Making

Weather conditions are critical in various aspects of daily life, particularly in transportation safety, where adverse weather can lead to accidents and disruptions. Traditional weather monitoring systems often rely on expensive sensors or human observation, which are prone to errors and may not provide real-time data necessary for immediate decision-making. To address this, an automated weather classification system is needed to identify weather conditions from single images accurately. This research proposes a deep learning approach leveraging transfer learning with the pre-trained InceptionV3 model to classify five distinct weather conditions: cloudy, foggy, rainy, sunny, and sunrise. A dataset was used to train the model of labelled weather images and fine-tuned through transfer learning to improve performance. The results show that the proposed transfer learning model achieved a training accuracy of 99.67% and a testing accuracy of 91.83%, outperforming other models. Additionally, the transfer learning model demonstrated superior generalization, with significantly lower testing loss than CNN.

I. Introduction

Rain and snowfall increase noise, modify point cloud resolution, and make traffic object detection difficult. Object segmentation under snowy conditions, 3-D light detection along roadsides, and backdrop point cloud extraction using the range. Object point clouds with different beam densities filter snowfall-related noise in the non-background point cloud [1]. Driving conditions, traffic infrastructure, and operating plans are all negatively impacted by inclement weather, especially heavy rain and the resulting wet road surface. Continuous real-time monitoring with high geographic granularity is necessary to avoid the potentially damaging consequences of stormy weather [2]. All kinds of weather should not be an issue for autonomous vehicles. An essential part of the autonomous cars perception stack is the ability to identify drivable road regions. While road region detection works effectively in nice weather, it fails miserably when the skies are overcast. To effectively identify drivable road zones in every weather condition, the latest deep learning technology to create a novel multimodal model is based on cameras and automotive radar [3]. Vehicles that can drive themselves and devices that help drivers rely on road scene analysis. Autonomous vehicles nowadays can see well in favourable weather, but they still have a way to go before they can handle situations where visibility is impaired. According to some, optimal road scene analysis can be achieved by combining classical vision with non-traditional sensors like infrared or lidar [4]. Event cameras have several uses in traffic flow detection because they are sensitive to moving targets but insensitive to stationary ones.