Cyclone Prediction from Remote Sensing Images Using Hybrid Deep Learning Approach Based on AlexNet

With the world feeling the negative effects of climate change, detecting and predicting severe weather occurrences is now an extremely vital and difficult task. Cyclones, a type of extreme weather phenomenon, have increased in frequency and severity in Indian subcontinent regions over the past few years. It is estimated that around three cyclones struck the east coastal region of India, causing substantial damage to people, farms, and infrastructure. Predicting cyclones ahead of time is crucial for avoiding or significantly lowering the devastating effects. The traditional methodologies employed numerical equations that demand strong experience and greater skills to obtain satisfactory prediction accuracy. Problems with domain expertise and the probability of human mistakes can be avoided with the help of Deep Learning (DL). As a result, in this work, we sought to forecast cyclone intensity using a Convolution Neural Network (CNN), a basic DL structure. To increase the CNN model's architecture and effectiveness, hybrid models such as Convolution Neural Network & Long short-term memory (CNN-LSTM) and AlexNet & Gated recurrent units (AlexNet-GRU) are developed. Data from the INSAT 3D satellite was utilized to develop and evaluate the DL model. We processed both the training and testing dataset and increase the training dataset using augmentation. All three DL models are tested and compared, the AlexNet-GRU model outperforms on the test data, with a relatively high accuracy of 93.35% and a low mean square error (MSE) of 215.