Multiple Time-step Predictive Models for Hurricanes in the North Atlantic Basin Based on Machine Learning Algorithms
The cost of damage caused by hurricanes in 2017 is estimated to be over 200 billion dollars. Quick and accurate prediction of the path of a hurricane and its strength would be very valuable in alleviating these losses. Machine learning based prediction models, in contrast to models based on physics, have been developed successfully in many problem domains. A machine learning system infers the modeling function from a training dataset. This project developed machine learning based prediction models to forecast the path and strength of hurricanes in the North Atlantic basin. Feature analysis was performed on the HURDAT2 dataset, which contains paths and strengths of past hurricanes. Artificial Neural Networks (ANNs) and Generalized Linear Model (GLM) approaches such as Tikhonov regularization were investigated to develop nine hurricane prediction models. Prediction accuracy of these models was compared using a testing dataset, disjoint from the training dataset. The coefficient of determination and the mean squared error were used as performance metrics. Post-processing metrics, such as geodesic error in path prediction and the mean wind speed error, were also used to compare different models. TLS linear regression model performed the best of out the nine models for one and two time steps, while the ANNs made more accurate predictions for longer periods. All models predicted location and strength with greater than .95 coefficient of determination for up to two days. My models predicted hurricane path in under a second with accuracy comparable to that of current models.
Satellite Modeling of Wildfire Susceptibility in California Using Artificial Neural Networking
Wildfires have become increasingly frequent and severe due to global climatic change, demanding improved methodologies for wildfire modeling. Traditionally, wildfire severities are assessed through post-event, in-situ measurements. However, developing a reliable wildfire susceptibility model has been difficult due to failures in accounting for the dynamic components of wildfires (e.g. excessive winds). This study examined the feasibility of employing satellite observation technology in conjunction with artificial neural networking to devise a wildfire susceptibility modeling technique for two regions in California. Timeframes of investigation were July 16 to August 24, 2017, and June 25 to December 8, 2017, for the Detwiler and Salmon August Complex wildfires, respectively. NASA’s MODIS imagery was utilized to compute NDVI (Normalized Difference Vegetation Index), NDWI (Normalized Difference Water Index), land surface temperature, net evapotranspiration, and elevation values. Neural network and linear regression modeling were then conducted between these variables and ∆NBR (Normalized Burn Ratio), a measure of wildfire burn severity. The neural network model generated from the Detwiler wildfire region was subsequently applied to the Salmon August Complex wildfire. Results suggest that a significant degree of variability in ∆NBR can be attributed to variation in the tested environmental factors. Neural networking also proved to be significantly superior in modeling accuracy as compared to the linear regression. Furthermore, the neural network model generated from the Detwiler data predicted ∆NBR for the Salmon August Complex with high accuracy, suggesting that if fires share similar environmental conditions, one fire’s model can be applied to others without the need for localized training.