Weather Image Classification Using a Modified CNN with EfficientNetB3

Identifying weather conditions from images plays a crucial role in computer vision and environmental information systems, particularly due to its practical applications in areas such as climate monitoring, intelligent transportation, and emergency disaster response.This study introduces an innovative Convolutional Neural Network (CNN) architecture built upon EfficientNetB3 for multi-class weather image classification.In addition to leveraging EfficientNetB3 as the primary feature extractor, the proposed approach incorporates batch normalization, a fully connected layer, and dropout in the final classification layer to enhance training stability and mitigate overfitting.These advancements align with the rapid progress in artificial intelligence.The dataset encompasses eleven distinct weather conditions.including dew frost, hail, lightning, rain, rainbow, rime, sandstorm, and snow.Additionally, all images were standardized to a resolution of 224 × 224 pixels and were subsequently divided into training, testing, and validation subsets with a distribution ratio of 60:20:20.After running a series of experiments, the suggested new model almost perfectly performed with a training accuracy of 0.9984, testing accuracy of 0.9345, and validation accuracy of 0.9375.Additionally, the macro averages for precision, recall, and F1-score were all 0.94, indicating that the model was consistent and reliable across various weather categories.Individually, some weather events like dew hail, lightning, and rainbows have been recognized with very high accuracy, whereas frost glaze, rime, rain, and snow have exhibited comparatively lower recognition performance.This is mainly because Truth is the last group with very similar visual characteristics.The experimental results demonstrate that the adapted EfficientNetB3 architecture offers a reliable and resource-efficient approach to classifying weather images within environmental monitoring systems.