Tornadoes-loss

Tornadoes-loss is a predictive service for calculating the loss in US dollars from a tornado.

Demo Site

http://tornadoes.azurewebsites.net/

Introduction

This is a Machine Learning experiment, which implements Predictive Analytic service in GIS projects using Azure Machine Learning.

System Architecture

Data

The data have being downloaded from NOAA's National Weather Service Storm prediction center .

The csv with the data of the tornadoes from 1950-2016 can be found here

Machine Learning

The data have being uploaded, filtered and analyzed so they would be ready for training the model.

Azure Machine Learning

For this procedure was used the Azure Machine Learning platform (link)

In the figure below shows the whole procedure for training and eveluating the model.

Modeling

There have being used many Machine Learning Algorithms, but it end up in two:

Two-Class Logistic Regression
Two-Class Averaged Perceptron

Evaluation

The Roc Curve is over the random guess

The evaluation of the two training model are shown below

Notice, that the accuracy was around 85% and the precision 87%.

Cortana Analytics Gallery

The experiment is published in the Cortana Analytics Gallery and can be found in the above link:

Tornadoes cost experiment

Web Service

Then it was published as a service using the Azure Machine Learning Platform and the end point returned as a Swagger API endpoint.

Web Application

Last, but not least, it has being developed a ASP.NET Core application which:

Serves static html content
Exposes a Web API.

The web application can be found here:

http://tornadoes.azurewebsites.net/

GIS development

For the need of the projected it was used JS libraries like D3.js, which creates very quickly svg inteactive maps.

Case Study

Tornado loss perdiction analysis service is a case study for predicting the cost for more or less than $5,000,000.

The user can fill the parameters of the form and by submitting it, get the result of the loss than will occure.

There are two ways of querying the service:

By selecting one month at a time.
By selecting all the months and get an annual result.

For the first way, the client can discover if separately what will happen in a state. For the second way, the client can discover that is the period that is more probably to have loss more tha $5,000,000.

Below, are shown 3 examples with their results:

tornado loss more than $5,000,000.
tornado loss less than $5,000,000.
tornado loss, 12 months prediction.

Example over than $5,000,000

Example less than $5,000,000

Example 12 month prediction

License

The whole project is under the MIT License

Charmatzis/Tornadoes-loss