Discrete Migratory Bird Optimizer with Deep Learning Driven Cyclone Intensity Prediction on Remote Sensing Images

Tropical Cyclones (TCs) are extreme climatic conditions that can crucially disrupt human life. Heavy rainfall and resilient winds that follow these systems can result in severe consequences for property and hamper social and economic growth in respective areas. Thus, accurate assessments of TC intensity is paramount for practical applications and theoretical research in predicting and preventing disasters. Satellite Cloud Images (SCIs) are a primary preferable and effective data source for the study of TCs. Efficient and accurate estimation of TC intensity is often challenging despite the remarkable success in different SCI-based studies. Recently, Machine Learning (ML) and Deep Learning (DL) methods have shown significant potential and gained fast development against big data, especially with images. Considerable progress has been made in applying Convolutional Neural Networks (CNNs) to predict and evaluate the intensity of TCs. This study focuses on developing a Discrete Migratory Bird Optimizer with Deep Learning Dirven Cyclone Intensity Prediction (DMBODL-CIP) technique on remote sensing images to estimate the intensity levels of TCs. To accomplish this, the DMBODL-CIP technique initially undergoes preprocessing in two phases: Bilateral Filtering (BF) and Adaptive Histogram Equalization (AHE)-based noise removal and contrast enhancement. The DMBODL-CIP technique utilizes a deep CNN-based SqueezeNet model for the feature extraction process. Then, a Deep Belief Network (DBN) model is used to predict TC intensity. Finally, the DMBO technique is employed for optimal hyperparameter selection of the DBN model, which assists in improving the overall prediction results. The proposed DMBODL-CIP approach was evaluated on a cyclone image dataset and a comparison study showed an RMSE of 6.02 kt outperforming existing techniques.