Integrating Satellite Data and Machine Learning for Rapid Forecasting of Severe Weather Events

Abstract

The large human scope of loss, property damage, and ecosystem disruptions associated with the frequency and severity of adverse weather changes are also increasing as a result of climate change. Existing weather forecasting infrastructures are relatively accurate but have a delayed response time in convergence and poor precision of forecasts of short-lived weather events like hurricanes, tornados, and flash floods. This paper discusses how to incorporate satellite based observational data with advanced machine learning (ML) algorithms to provide faster and more precise anticipation of severe weather events. Combining the features of multispectral, thermal, and radar-based satellite imagery, ML can locate obscured spatiotemporal patterns and deliver short-range predictions with extended lead times. A combination of convolutional neural networks (CNNs) to identify spatial features, and recurrent neural networks (RNNs) to detect the trends in time was used. The outcomes show that theML-forecasting system has a higher accuracy in determining the intensity of storms and in tracking the path of the storm than the traditional numerical weather prediction (NWP) models. Notwithstanding, the real-life application is still subject to practical limitations, including latency of the data, loss of information through clouds, overfitting of models, and modeling costs in real-time application. Future work needs to be on the integration of multiple satellite constellations, transfer learning to make the models more general and cloud-based systems that can deliver usable preliminary warnings in advance of a disaster threatening a target region.