Correlation - Hurricane analysis, part 3/3¶
Table of Contents¶
- Access aggregated hurricane data
- Manage missing sensor data
- Does intensity of hurricanes increase over time?
- Does the number of hurricanes increase over time?
- Does hurricane wind speed increase over time?
- Analyzing hurricane category over time by basin
- Analyzing hurricane wind speed over time by basin
- Does eye pressure decrease over time?
- Do hurricanes linger longer over time?
- Do hurricanes travel longer inland over time?
- Correlate observations over time
- Conclusion
- References
This is the third part to a three part set of notebooks that process and analyze historic hurricane tracks. In the previous notebooks we saw:
Part 1
- Downloading historic hurricane datasets using Python
- Cleaning and merging hurricane observations using DASK
- Aggregating point observations into hurricane tracks using ArcGIS GeoAnalytics server
Part 2
- Visualize the locations of hurricane tracks
- Different basins and the number of hurricanes per basin
- Number of hurricanes over time
- Seasonality in occurrence of hurricanes
- Places where hurricanes make landfall and the people are affected
In this notebook you will analyze the aggregated tracks to answer important questions about hurricane severity and how they correlate over time.
Import the libraries necessary for this notebook.
# import ArcGIS Libraries
from arcgis.gis import GIS
from arcgis.geometry import filters
from arcgis.geocoding import geocode
from arcgis.features.manage_data import overlay_layers
from arcgis.geoenrichment import enrich
# import Pandas for data exploration
import pandas as pd
import numpy as np
from scipy import stats
# import plotting libraries
import matplotlib.pyplot as plt
import seaborn as sns
%matplotlib inline
# import display tools
from pprint import pprint
from IPython.display import display
# import system libs
from sys import getsizeof
# Miscellaneous imports
import warnings
warnings.filterwarnings('ignore')
gis = GIS(url='https://pythonapi.playground.esri.com/portal', username='arcgis_python', password='amazing_arcgis_123')
Access aggregated hurricane data¶
Below, we access the tracks aggregated using GeoAnalytics in the previous notebook.
hurricane_tracks_item = gis.content.search('title:hurricane_tracks_aggregated_ga', item_type="feature layer")[0]
hurricane_fl = hurricane_tracks_item.layers[0]
The GeoAnalytics step calculated summary statistics of all numeric fields. However, only a few of the columns are of interest to us.
pprint([f['name'] for f in hurricane_fl.properties.fields], compact=True, width=80)
Below we select the following fields for the rest of this analysis.
fields_to_query = ['objectid', 'count', 'min_season', 'any_basin', 'any_sub_basin',
'any_name', 'mean_latitude_merged', 'mean_longitude_merged',
'max_wind_merged', 'range_wind_merged', 'min_pressure_merged',
'range_pressure_merged', 'max_eye_dia_merged', 'track_duration',
'end_datetime', 'start_datetime']
Query hurricane tracks into a Spatially Enabled DataFrame
¶
all_hurricanes_df = hurricane_fl.query(out_fields=','.join(fields_to_query), as_df=True)
all_hurricanes_df.shape
There are 12,362
hurricanes identified by GeoAnalytics aggregate tracks tool. To get an idea about this aggregated dataset, call the head()
method.
all_hurricanes_df.head()
To better analyze this data set, the date columns need to be changed to a format that Pandas understands better. This is accomplished by calling to_datetime()
method and passing the appropriate time columns.
all_hurricanes_df['start_datetime'] = pd.to_datetime(all_hurricanes_df['start_datetime'])
all_hurricanes_df['end_datetime'] = pd.to_datetime(all_hurricanes_df['end_datetime'])
all_hurricanes_df.index = all_hurricanes_df['start_datetime']
all_hurricanes_df.head()
The track duration and length columns need to be projected to units (days, hours, miles) that are meaningful for analysis.
all_hurricanes_df['track_duration_hrs'] = all_hurricanes_df['track_duration'] / 3600000
all_hurricanes_df['track_duration_days'] = all_hurricanes_df['track_duration'] / (3600000*24)
Query landfall tracks layer into a Spatially Enabled DataFrame
¶
We query the landfall tracks layer created in the pervious notebook into a DataFrame.
inland_tracks = gis.content.search('hurricane_landfall_tracks')[0]
fields_to_query = ['min_season', 'any_basin','any_name', 'max_wind_merged',
'min_pressure_merged', 'track_duration','end_datetime',
'start_datetime', 'analysislength']
landfall_tracks_fl = inland_tracks.layers[0]
landfall_tracks_df = landfall_tracks_fl.query(out_fields=fields_to_query).df
landfall_tracks_df.head(3)
Manage missing sensor data¶
Before we can analyze if hurricanes intensify over time, we need to identify and account for missing values in our data. Sensor measurements such as wind speed, atmospheric pressure, eye diameter, generally suffer from missing values and outliers. The reconstruct tracks tool has identified 12,362
individual hurricanes that occurred during the past 169
years.
all_hurricanes_df.shape
Visualize missing records¶
An easy way to visualize missing records is to hack the heatmap
of seaborn
library to display missing records. The snippet below shows missing records in yellow color.
missing_data_viz = all_hurricanes_df.replace(0,np.NaN)
missing_data_viz = missing_data_viz.replace(-9999.0,np.NaN)
missing_data_viz['min_pressure_merged'] = missing_data_viz['min_pressure_merged'].replace(100.0,np.NaN)
plt.figure(figsize=(10,10))
missing_data_ax = sns.heatmap(missing_data_viz[['max_wind_merged', 'min_pressure_merged',
'max_eye_dia_merged', 'track_duration']].isnull(),
cbar=False, cmap='viridis')
missing_data_ax.set_ylabel('Years')
missing_data_ax.set_title('Missing values (yellow) visualized as a heatmap')
All three observation columns - wind speed, atmospheric pressure and eye diameter, suffer from missing values. In general as technology improved over time, we were able to collect better data with fewer missing observations. In the sections below, we attempt to fill these values using different techniques. We will compare how they fare and pick one of them for rest of the analysis.
Missing value imputation¶
Technique 1: Drop missing values: An easy way to deal with missing values is to drop those record from analysis. If we were to do that, we lose over a third of the hurricanes.
hurricanes_nona = missing_data_viz.dropna(subset=['max_wind_merged','min_pressure_merged'])
hurricanes_nona.shape
Technique 2: Fill using median value: A common technique is to fill using median value (or a different measure of centrality). This technique computes the median of the entire column and applies that to all the missing values.
fill_values = {'max_wind_merged': missing_data_viz['max_wind_merged'].median(),
'min_pressure_merged': missing_data_viz['min_pressure_merged'].median(),
'track_duration_hrs': missing_data_viz['track_duration_hrs'].median()}
hurricanes_fillna = missing_data_viz.fillna(value=fill_values)
Technique 3: Fill by interpolating between existing values: A sophisticated approach is to interploate a missing value based on two of its closest observations.
hurricanes_ipl = missing_data_viz
hurricanes_ipl['max_wind_merged'] = hurricanes_ipl['max_wind_merged'].interpolate()
hurricanes_ipl['min_pressure_merged'] = hurricanes_ipl['min_pressure_merged'].interpolate()
hurricanes_ipl['track_duration_hrs'] = hurricanes_ipl['track_duration_hrs'].interpolate()
Visualize all 3 techniques
To compare how each of these techniques fared, we will plot the histogram of wind speed column after managing for missing values.
fig, ax = plt.subplots(1,3, sharex=True, figsize=(15,5))
fig.suptitle('Comparing effects of missing value imputations on Wind speed column',
fontsize=15)
hurricanes_nona['max_wind_merged'].plot(kind='hist', ax=ax[0], bins=35, title='Drop null values')
hurricanes_fillna['max_wind_merged'].plot(kind='hist', ax=ax[1], bins=35, title='Impute with median')
hurricanes_ipl['max_wind_merged'].plot(kind='hist', ax=ax[2], bins=35, title='Impute via interpolation')
for a in ax:
a.set_xlabel('Wind Speed')
Next, we will plot the histogram of atmospheric pressure column after managing for missing values.
fig, ax = plt.subplots(1,3, sharex=True, figsize=(15,5))
fig.suptitle('Comparing effects of missing value imputations on Pressure column',
fontsize=15)
hurricanes_nona['min_pressure_merged'].plot(kind='hist', ax=ax[0], title='Drop null values')
hurricanes_fillna['min_pressure_merged'].plot(kind='hist', ax=ax[1], title='Impute with median')
hurricanes_ipl['min_pressure_merged'].plot(kind='hist', ax=ax[2], title='Impute via interpolation')
for a in ax:
a.set_xlabel('Atmospheric Pressure')
Fill using interpolation preserves shape of the original distribution. So it will be used for further anlaysis.
Does intensity of hurricanes increase over time?¶
This last part of this study analyzes if there a temporal trend in the intensity of hurricanes. A number of studies have concluded that anthropogenic influences in the form of global climate change make hurricanes worse and dangerous. We analyze if such patterns can be noticed from an empirical standpoint.
Does the number of hurricanes increase over time?¶
ax = all_hurricanes_df['min_season'].hist(bins=50)
ax.set_title('Number of hurricanes per season')
From the previous notebook, we noticed the number of hurricanes recorded has been steadily increasing, partly due to advancements in technology. We notice a reduction in number of hurricanes after 1970s. Let us split this up by basin and observe if the trend is similar.
fgrid = sns.FacetGrid(data=all_hurricanes_df, col='any_basin', col_wrap=3,
sharex=False, sharey=False)
fgrid.map(plt.hist, 'min_season', bins=50)
fgrid.set_axis_labels(x_var='Seasons', y_var='Frequency of hurricanes')
Plotting the frequency of hurricanes by basin shows a similar trend with the number of hurricanes reducing globally after 1970s. This is consistent with certain studies (1). However, this is only one part of the story. Below, we continue to analyze if the nature of hurricanes itself is changing, while the total number may reduce.
Does hurricane wind speed increase over time?¶
To understand if wind speed increases over time, we create a scatter plot of min_season
against the max_wind_merged
column. The seaborn
plotting library can enhance this plot with a correlation coefficient and its level of signifance (p value).
jgrid = sns.jointplot(x='min_season', y='max_wind_merged', data=hurricanes_ipl,
kind='reg', joint_kws={'line_kws':{'color':'green'}}, height=7, space=0.5)
j = jgrid.annotate(stats.pearsonr)
j = jgrid.ax_joint.set_title('Does hurricane wind speed increase over time?')
From the plot above, we notice a small positive correlation. Wind speeds are observed to increase with time. The small p-value
suggests this correlation (albeit small) is statistically significant. The plot above considers hurricanes across all the basins and regresses that against time. To get a finer picture, we need to split the data by basins and observe the correlation.
# since there are not many hurricanes observed over South Atlantic basin (SA),
# we drop it from analysis
hurricanes_major_basins_df = hurricanes_ipl[hurricanes_ipl['any_basin'].isin(
['WP','SI','NA','EP','NI','SP'])]
Define a function that can compute pearson-r
correlation coefficient for any two columns across all basins.
def correlate_by_basin(column_a, column_b, df=hurricanes_major_basins_df,
category_column = 'any_basin'):
# clean data by dropping any NaN values
df_nona = df.dropna(subset=[column_a, column_b])
# loop through the basins
basins = list(df[category_column].unique())
return_dict = {}
for basin in basins:
subset_df = df_nona[df_nona[category_column] == basin]
r, p = stats.pearsonr(list(subset_df[column_a]), list(subset_df[column_b]))
return_dict[basin] = [round(r,4), round(p,4)]
# return correlation coefficient and p-value for each basin as a DataFrame
return_df = pd.DataFrame(return_dict).T
return_df.columns=['pearson-r','p-value']
return return_df
Analyzing hurricane wind speed over time by basin¶
Below we plot a grid of scatter plots with linear regression plots overlaid over them. The seaborn
library's lmplot()
function makes it trivial to accomplish this in a single command.
fgrid = sns.lmplot('min_season', 'max_wind_merged', col='any_basin',
data=hurricanes_major_basins_df, col_wrap=3,
sharex=False, sharey=False, line_kws={'color':'green'})
From the scatter plots above, we notice the wind speeds in most basins show a slight positive trend, with North Atlantic being an exception. To explore this further, we compute the correlation coefficient and its p-value below.
wind_vs_season = correlate_by_basin('min_season','max_wind_merged')
print('Correlation coefficients for min_season vs max_wind_merged')
wind_vs_season
The table above displays the correlation coefficient of hurricane wind speed over time. Hurricanes over the Southern Pacific basin exhibit a positive trend of increasing wind speeds. The r
value over the North Atlantic shows a weak negative trend. Since all the p-value
s are less than 0.05
, these correlations are statistically significant.
Analyzing hurricane category over time by basin¶
Hurricanes are classified on a Saffir-Simpson scale of 1-5
based on their wind speed. Let us compute this column on the dataset and observe if there are temporal aspects to it.
def categorize_hurricanes(row, wind_speed_column='max_wind_merged'):
wind_speed = row[wind_speed_column] * 1.152 # knots to mph
if 74 <= wind_speed <= 95:
return '1'
elif 96 <= wind_speed <= 110:
return '2'
elif 111 <= wind_speed <= 129:
return '3'
elif 130 <= wind_speed <= 156:
return '4'
elif 157 <= wind_speed <= 500:
return '5'
hurricanes_major_basins_df['category_str'] = hurricanes_major_basins_df.apply(categorize_hurricanes,
axis=1)
hurricanes_major_basins_df['category'] = pd.to_numeric(arg=hurricanes_major_basins_df['category_str'],
errors='coerce', downcast='integer')
hurricanes_major_basins_df.head(2)
We will create violin and bar plots to visualize the number of hurricane categories over different basins.
fig, ax = plt.subplots(1,2, figsize=(15,6))
vplot = sns.violinplot(x='any_basin', y='category', data=hurricanes_major_basins_df, ax=ax[0])
vplot.set_title('Number of hurricanes per category in each basin')
cplot = sns.countplot(x='any_basin', hue='category_str', data=hurricanes_major_basins_df,
hue_order=['1','2','3','4','5'], ax=ax[1])
cplot.set_title('Number of hurricanes per category in each basin')
We notice all basins are capable of generating major hurricanes (over 3). The Eastern Pacific basin appears to have a larger than the proportional number of major hurricanes. Below, we will regress the hurricane category against time to observe if there is a positive trend.
kde_regplot = sns.jointplot(x='min_season', y='category',
data=hurricanes_major_basins_df, kind='kde',
stat_func=stats.pearsonr).plot_joint(sns.regplot,
scatter=False)
kde_regplot.ax_joint.set_title('Scatter plot of hurricane categories over seasons')
Even at a global level, we notice a strong positive correlation between hurricane category and seasons. Below, we will split this across basins to observe if the trend holds well.
wgrid = sns.lmplot('min_season', 'category', col='any_basin',
data=hurricanes_major_basins_df, col_wrap=3,
sharex=False, sharey=False, line_kws={'color':'green'})
category_corr_df = correlate_by_basin('min_season','category', df=hurricanes_major_basins_df)
print('Correlation coefficients for min_season vs hurricane category')
category_corr_df
Thus, at both global and basin scales, we notice a positive trend in the number of hurricanes of category 4
and higher, while there is a general reduction in the quantity of hurricanes per season. This is along the lines of several studies [1] [2] [3] [4]. Thus while the total number of hurricanes per season may reduce, we notice an increase in the intensity of them.
Does eye pressure decrease over time?¶
Just like a high wind speed, lower atmospheric pressure increases the intensity of hurricanes. To analyze this, we produce a scatter grid of min_pressure_merged
column and regress it against min_season
column. We split this by basins.
pgrid = sns.lmplot('min_season', 'min_pressure_merged', col='any_basin',
data=hurricanes_major_basins_df, col_wrap=3,
sharex=False, sharey=False, line_kws={'color':'green'})