Efficient Weather Classification Using DenseNet and EfficientNet

Abstrak

Classifying images of weather conditions using deep learning models is a challenging task due to the computational intensity and resource requirements. To deploy AI models on resource-constrained devices like smartphones and IoT devices, compact and computationally lightweight models are necessary. Efficient deep learning models for weather classification are essential to reduce energy consumption and costs, making AI more accessible and sustainable. To the best of our knowledge, there are limited studies comparing MobileNet, DenseNet, and EfficientNet as efficient models and did not report any hyperparameter optimization. Our study contributes by investigating efficient models with hyperparameter optimization. Firstly, we measured the inference speed of 14 models, namely MobileNet, MobileNetV2, MobileNetV3, EfficientNetB0, EfficientNetV2B0, NASNetMobile, DenseNet121, VGG16, Xception, InceptionV3, ResNet50, ResNet50V2, ConvNeXtTiny, and InceptionResNetV2. Then, the top-7 fast models, which are MobileNet, MobileNetV2, MobileNetV3, EfficientNetB0, EfficientNetV2B0, NASNetMobile, and DenseNet121, were benchmarked for their accuracy. The models were compared by a small dataset having four classes: cloudy, rain, shine, and sunrise. Batch size and learning rate for each model were optimized by grid search method. It turns out that DenseNet121 achieved the best and the most balanced validation and testing accuracy, 0.9821 and 0.9837, followed by EfficientNetB0 with 0.9821 and 0.9740 respectively. This study is important to find efficient models with optimal comparison.