Towards Automatic Weather Classification Using DCNNs

Abstract

The utilization of remote sensing (RS) technology has resulted in the extensive accessibility of a significant amount of satellite image data. In order to ensure the successful execution of the RS in real-life scenarios, it is imperative to create effective and adaptable solutions that can be utilized across different transdisciplinary domains. Deep Convolutional Neural Networks (CNNs) are frequently used to accomplish the goal of fast analysis and precise categorization in RS imaging. This study introduces a unique residual network known as ResNet101. The network comprises FC-1024 fully connected layers, dropout layers, a thick layer, and data augmentation algorithms. To resolve the issue of similarity between different classes, architectural enhancements are implemented. On the other hand, imbalanced classes are dealt with by employing data augmentation techniques. The ResNet101 model use the rigorous Large-Scale Cloud pictures Dataset for Meteorology Research (LSCIDMR), which has 10 classes and a multitude of highresolution photos. The goal of the model is to precisely classify these photos into their respective categories. The model we have created outperforms numerous previously published deep learning algorithms in terms of Precision, Accuracy, and F1 scores. The accuracy reaches up to 99% and approximately 92%, respectively.