Milestone 1:
Navigation: Introduction → Preliminary Work → Creating a Model Using Weka → Results and Implications
Milestone 1:
To begin creating our prediction equations, we looked at data from hurricanes that occurred between 1983 and 2005. A range of variables were included in the database:
Because such a large quantity of data was available to us, we started to graph the data in Matlab by individual storms in a specific year in order to see overall patterns and relationships between variables. With intensity, we looked at the relationship between time and intensity and the storm's intensity with the radius of the winds. A majority of the time, storms had a linear increase of intensity for the first part of the storm, before stagnating and then decreasingly linearly, like in Hurricane Danielle from 1986 (click on figure to right to enlarge). Thus, the previous intensity would have an impact on the next intensity. Many of the storms showed a moderately strong correlation between the radius of the winds and the intensity. Hurricane Frances, also in 1986, had an excellent correlation between intensity and 34 knot winds, as shown in the graphs (click on figure to left to enlarge).
(click on figure to right to enlarge).
We also compared latitude and longitude to the radius of the winds. Even though we did not formally calculate correlation values because of missing data, the graphs showed a fairly linear relationship between latitude and radius of 34 knot winds. An example of this relationship can be found in the graphs from 1985's Hurricane Isabel (click on figure below to enlarge). However, because the relationships would vary wildly between storms and years, we hypothesized that the radius of the winds would not be the most significant factor in determining the next three hours of change in a storm.
The following is a video demonstration of the milestone:
Milestone 2:
After displaying the variables on graphs to create a framework of variables that appeared to have a relation, we plotted the storm trajectory over the length of the storm. This would give us a visual interpretation of intensity versus geographical location. As well, we would be able to see how position over land or sea would affect the hurricane’s intensity. By adding in sea surface temperatures, we would also look at their effect on the categorizing of the hurricane. As shown in the figure from 1985 (click on figure to right to enlarge), the storms began to die down once they started traveling over land. As the storms moved to the north and east, there was also a trend to decrease in intensity. These trends gave us some idea of what we should see in a statistical model that accurately predicted changes in the hurricane.
Seeing the entire track of the storm also showed that storms would stay roughly in the same direction as they moved. It seemed that the previous direction would most likely determine the direction the storm would head in the next three hours, with steering provided by the relationships we had seen earlier.
The following is a video demonstration of the milestone:
Milestone 3:
Now that we had a general idea of trends to look for and incorporate into our models, we began to prepare our data for analysis in Weka. Based on the trends we had seen in our graphs as well as previous knowledge on hurricane formation, we compile a list of variables that we felt would be the most helpful, including: