Pre-Trained Convolutional Neural Network Benchmark Untuk Pemodelan Cuaca Multi-Class

Pre-Trained Convolutional Neural Network Benchmark For Multi-Class Weather Modeling

Kompleksitas pola cuaca yang sulit diprediksi menimbulkan tantangan dalam pengembangan sistem klasifikasi otomatis berbasis citra. Penelitian ini membandingkan kinerja arsitektur Convolutional Neural Network (CNN) pra-latih, yaitu ResNet50, VGG16, AlexNet, dan InceptionV3, untuk klasifikasi citra cuaca multi-kelas menggunakan dataset Kaggle yang berisi 860 citra dengan kategori Cloudy, Rain, Shine, dan Sunrise. Evaluasi dilakukan dengan variasi rasio pembagian data (50:50, 60:40, 70:30, 80:20, 90:10) serta pengaturan hyperparameter seperti learning rate, batch size, dan epoch, menggunakan metrik akurasi, presisi, recall, dan F1-score. Hasil menunjukkan bahwa InceptionV3 memiliki performa terbaik dengan akurasi pelatihan 98% dan validasi 96% pada rasio 70:30, diikuti ResNet50 dengan akurasi validasi hingga 94%. AlexNet dan VGG16 menunjukkan performa lebih rendah, dengan VGG16 memiliki akurasi terendah. Temuan ini memberikan wawasan penting bagi pengembangan sistem prakiraan cuaca berbasis citra yang lebih akurat dan efisien.

The complexity of unpredictable weather patterns presents challenges in developing automated image-based classification systems. This study compares the performance of pre-trained Convolutional Neural Network (CNN) architectures—ResNet50, VGG16, AlexNet, and InceptionV3 for multi-class weather image classification using a Kaggle dataset containing 860 images categorized into Cloudy, Rain, Shine, and Sunrise. Evaluation was conducted with varying data split ratios (50:50, 60:40, 70:30, 80:20, 90:10) and hyperparameter settings such as learning rate, batch size, and epochs, using accuracy, precision, recall, and F1-score as performance metrics. The results indicate that InceptionV3 achieved the best performance, with a training accuracy of 98% and validation accuracy of 96% at a 70:30 split ratio, followed by ResNet50 with validation accuracy up to 94%. AlexNet and VGG16 showed lower performance, with VGG16 recording the lowest accuracy. These findings provide valuable insights for developing more accurate and efficient image-based weather forecasting systems.