Cloud And Rain Streak Image Analysis For Rainfall Prediction Using Deep Learning Approaches

Rainfall prediction plays a pivotal role in weather forecasting and environmental management. It has significant implications for agriculture, water resource management, disaster preparedness, and many other sectors. The use of deep learning techniques for predicting rainfall based on cloud image and rain streak image analysis has been gaining more and more importance lately. The study highlights the importance of having accurate rainfall forecasts and introduces a new data preprocessing methodology that aims to improve the precision of rainfall predictions. Day/Night discrimination based on HSV (Hue, Saturation, Value), resizing images to a standardized dimension of $128 \times 128$, and introducing random image rotations contribute to the robustness of the dataset used for training deep learning models. Effective preprocessing is essential for enabling accurate rainfall predictions by helping models learn and generalize effectively from visual cues in cloud images. The study evaluated the performance of five deep learning architectures for rainfall prediction using cloud image analysis: CNN, RNN (LSTM), DNN, DenseNet-169, and ResNet-50. To train and test the deep learning models, rainfall data from four publicly available datasets, namely Hindustan Times and The Weather Channel, Swimcat, and Unsplash, was used. The DenseNet-169 architecture achieved the best accuracy of 98.5%, followed by CNN at 94.5%. DNN architecture had an accuracy of $\mathbf{8 4 \%}$, followed by the ResNet-50 architecture with an accuracy of 77%, and RNN architecture with an accuracy of 75%. These results suggest that DenseNet-169 is the most promising architecture for rainfall prediction using cloud image analysis. However, further research is needed to improve the accuracy of RNNs and other architectures for this task.