A LIGHTWEIGHT DUAL-BRANCH CNN–MOBILENETV2 ENSEMBLE FOR MULTI-CLASS WEATHER IMAGE CLASSIFICATION

Abstract

For outdoor autonomous systems that must operate safely and reliably to identifying the prevailing weather directly from a captured Image has become a practical alternative to dedicated sensing hardware. A recurring difficulty is that a compact Convolutional Neural Network (CNN) trained only on the target data learns cues specific to that data yet fails to generalize when the training set is small, whereas a transfer learning model such as VGG16 generalizes more reliably but carries a parameter budget too large for on-device use. This paper introduces a two-branch architecture that runs a small custom CNN in parallel with a fine-tuned MobileNetV2 and merges its penultimate feature vectors before a shared softmax classifier that combines the data-specific and ImageNet-transferred evidence within the network rather than by choosing one over the other. The method is benchmarked against different deep learning models such as a custom CNN, a simple Recurrent Neural Network (RNN), VGG16, and a standalone MobileNetV2 on a four-category weather image collection covering cloudy, rainy, shine, and sunrise scenes. The proposed design attains a test accuracy of 0.952 together with a macro F1-score of 0.963, surpassing the best single reference (VGG16 at 0.925) and its own CNN branch by a clear margin, while keeping a MobileNet-scale footprint. An experimental analysis further shows that fusion yields the tightest spread between training and test accuracy among all configurations that indicating the advantage stems from better generalization rather than added capacity.