This repository contains codes that demonstrate the use of Project AEDES for data collection on remote sensing using LANDSAT, MODIS and SENTINEL. Full repository is linked here.

Author: Xavier Puspus
Affiliation: Cirrolytix Research Services

Install using:

foo@bar:~$ pip install aedes

Import the modules of the package using:

import aedes
from aedes.remote_sensing_utils import generate_random_ee_points, df_to_ee_points, get_satellite_measures_from_points
from aedes.remote_sensing_utils import visualize_on_map

from aedes.osm_utils import reverse_geocode_points

from aedes.automl_utils import perform_clustering, perform_classification

This packages uses Google Earth Engine (sign-up for access here) to query remote sensing data. To authenticate, simply use:

aedes.remote_sensing_utils.authenticate()

This script will open a google authenticator that uses your email (provided you've signed up earlier) to authenticate your script to query remote sensing data. After authentication, initialize access using:

aedes.remote_sensing_utils.initialize()

First, find the bounding box geojson of an Area of Interest (AOI) of your choice using this link.

Use the one-liner code get_satellite_measures_from_points to extract NDVI, NDWI, NDBI, Aerosol Index (Air Quality), Surface Temperature, Precipitation Rate and Relative Humidity for your preset number of points of interest sample_points within a specified date duration date_from to date_to.

%%time
QC_AOI = [[[120.98976275,14.58936896],
           [121.13383232,14.58936896],
           [121.13383232,14.77641364],
           [120.98976275,14.77641364],
           [120.98976275,14.58936896]]] # Quezon city
           
# Get random points sampled from the AOI if you don't have ground truth data points yet.
# You can also generate your own Earth Engine Points from your own long-lat pairs using generate_random_ee_points()
points = generate_random_ee_points(QC_AOI, sample_points=50)

# Get satellite features on each point
qc_df = get_satellite_measures_from_points(points, QC_AOI, 
                                              date_from='2017-07-01', 
                                              date_to='2017-09-30')

This package also provides an easy-to-use one-liner reverse geocoder that uses Nominatim

%%time
rev_geocode_qc_df = reverse_geocode_points(qc_df)
rev_geocode_qc_df.head()

This packages uses KMeans as the unsupervised learning technique of choice to perform clustering on the geospatial data enriched with normalized indices, air quality and surface temperatures with your chosen number of clusters.

clustering_model = perform_clustering(rev_geocode_qc_df, n_clusters=3)
rev_geocode_qc_df['labels'] = pd.Series(clustering_model.labels_)

This packages also provides the capability of visualizing all the points of interest with their proper labels using one line of code.

vizo = visualize_on_map(rev_geocode_qc_df)
vizo

The package needed is imported as follows:

from aedes.osm_utils import initialize_OSM_network, get_OSM_network_data 

In order to initialize and create an OpenStreetMap (OSM) network from a geojson of an AOI, use:

%%time
network = initialize_OSM_network(aoi_geojson)

In order to pull data for, say, healthcare facilities (more documentation on amenities here), use this one-liner:

final_df, amenities_df, count_distance_df = get_OSM_network_data(network,
                     satellite_df,
                     aoi_geojson,
                    ['clinic', 'hospital', 'doctors'],
                    5,
                    5000,
                    show_viz=True)

This function pulls the count and distance of each node from a possible healthcare facility (for this example). It also outputs the original dataframe concatenated with the count and distances. The actual amenities data is also returned. We can then pass the resulting final_df dataframe into another clustering algorithm to produce dengue risk clusters with the added health capacity features.

To query for Google search trends, import:

from aedes.social_listening_utils import get_search_trends

Find the specific ISO tag for the location (requirement by Google) here, then use:

iso_geotag = "PH-00"
search_df = get_search_trends(iso_geotag)

This pulls data for the top 5 dengue-related searches within a geolocation dictated by an ISO tag listed and described here. Below is a sample:

In order to demonstrate the functionalities of using the AEDES python package, we can use Streamlit to display a web application that takes in a geojson and outputs the hotspots and the recommended cities at risk. Clone this repository, cd into it and follow the instructions below.

Install streamlit using:

foo@bar:~$ pip install streamlit

Run streamlit hello to see if the installation was successful.

Simply run the code below to run a local version of your web application that outputs the at-risk areas as hotspots on a map as well as a subsequent list of places to prioritize disease-related proactive measures.

The one below is for a dengue hotspot map for Quezon City, Philippines.

This other screenshot shows the web application demonstrating the use of the geospatial modelling in outputting locations of high-risk areas.

Another example for Cotabato City, Philippines is shown below.

We have also added functionality to this package that performs tree-based pipeline optimization (TPOT) that optimizes machine learning pipelines using genetic algorithm as described here.

Data preparation is still required as we as train-test splitting as needed. Using the one-liner perform_classification or perform_regression, we can train a machine learning model hundreds to thousands of times in different configurations, feature engineering methodologies and various models in order to output:

• the best ml model (in-memory and saved as a pickle file, default is best_model.pkl)
• the best ml pipeline which is a python script that describes the ml methodology
• and the feature importances (both as a dataframe, and as a plot)

model, feature_imps_df = perform_classification(X_train, y_train)

The output should look like this:

which also generates a python script of the best machine learning model pipeline similar to this script.

We can use perform_clustering() to automatically create 5 clusters from the input data, with features preselected based on the INFORM risk framework. The packages needed are:

from aedes.automl_utils import perform_clustering
import pandas as pd

To demonstrate one example, say for Lack of Coping Capacity, the features needed are ['longitude', 'latitude', 'nearest_clinic_hospital_doctors_1', 'nearest_clinic_hospital_doctors_2', 'nearest_clinic_hospital_doctors_3', 'nearest_clinic_hospital_doctors_4', 'nearest_clinic_hospital_doctors_5', 'count_clinic_hospital_doctors_within_5.0km']. As long as the input data contains the same columns, we can just perform clustering as follows:

df = pd.read_csv("flatfile_of_your_dataset.csv")

loc_features = ['longitude',
 'latitude',
 'nearest_clinic_hospital_doctors_1',
 'nearest_clinic_hospital_doctors_2',
 'nearest_clinic_hospital_doctors_3',
 'nearest_clinic_hospital_doctors_4',
 'nearest_clinic_hospital_doctors_5',
 'count_clinic_hospital_doctors_within_5.0km']
 
 # Perform clustering (this model can be saved and re-loaded later using joblib)
loc_model = perform_clustering(df, 
                               features=loc_features,
                               n_clusters=5)
# Create the labels                               
loc_labels = pd.Series(loc_model.labels_)

# Create INFORM risk dataframe for Lack of Coping Capacity
loc_full_df = df[loc_features].drop_duplicates()
loc_full_df['loc_labels'] = loc_labels

We then perform analysis on categorical comparison of feature distribution as demonstrated by the sample images below:

Nearest Clinics/Hospitals	Count of Clinics/Hospitals

We see from above that, in decreasing count of and increasing distance from coping facilities, the cluster labels are: [2, 0, 3, 4, 1]. We can then create the Lack of Coping Capacity risk labels by re-labelling the clusters in this order and visualize the risk map as follows (darker is riskier):

from aedes.remote_sensing_utils import visualize_on_map

clusters_to_loc_risk_df = pd.DataFrame({"labels":list(range(1, 6)),
                                        "loc_labels":[2, 0, 3, 4, 1]})

points_df = loc_full_df.merge(clusters_to_loc_risk_df)

visualize_on_map(points_df)

We can also save and load the clustering model as follows:

import joblib

# Save the clustering model as a pickle file
joblib.dump(loc_model, "loc_clustering_model.pkl")

# Load the clustering model from a pickle file
loc_model = joblib.load("loc_clustering_model.pkl")