Detecting and mapping slums using open data: a case study in Kenya

Figures & data

Figure 1. Study areas in Kenya.

Figure 2. Methodology workflow.

Table 1. Candidate indicators for slums.

Candidate indicator group (number of indicators in group)	Data source	Description (Year/period)	Hypothesis	Processing steps
Class 1: Durable housing – UN Habitat requirement
Hazardous locations (1)	OpenStreetMap (2017)	Polyline vector (2014)	Slums are more likely to be in closer proximity to major roadways.	Primary and secondary roads from each data source were extracted and merged into a single dataset.
	Google Map Maker (2017)	Polyline vector (2014)
Image texture (48)	Landsat 8 panchromatic band – Annual median composite (Google Earth Engine 2017)	15 m spatial resolution (2014)	Certain texture measures at different scales (window sizes) will provide good separability between slums and other non-slum areas.	Various texture measures were extracted: contrast, correlation, dissimilarity, entropy, homogeneity, mean, second moment and variance. A unit shift value of 1 pixel in both x and y directions was used for these calculations. Each texture measure was extracted at different window sizes: 3 × 3, 5 × 5, 7 × 7, 9 × 9, 11 × 11 and 13 × 13. This resulted in a total of 48 image texture measures.
Raw image bands (9)	Landsat 8 – Image bands 1–7 and 10–11 Annual median composite (Google Earth Engine 2017)	Image bands 1 −7, 30 m spatial resolution (2014) Image bands 10–11, 100 m spatial resolution (2014)	Certain raw image bands are more sensitive to slums than others.	Nine raw image band indicators were computed.
Class 2: Adequate living space – UN Habitat requirement
Population density (1)	LandScan (ORNL 2017)	Raster image with 1 km by 1 km spatial resolution (2014)	Slums have higher population densities compared to other types of communities.	Population density was extracted from the dataset for each 1 km × 1 km grid cell.
Birth rates(1)	WorldPop birth rates (WorldPop 2017)	Raster image with 100 m by 100 m spatial resolution (2015)	Slums have higher birth rates compared to other types of communities.	Birth rates were extracted from the dataset for each 100 m × 100 m grid cell.
Class 3: Access to safe water – UN Habitat requirement
Water kiosks (1)	OpenStreetMap (2017)	Point vector (updated until 2014)	Slums have high a large number of water kiosks.	The XY location of water kiosks was extracted from both data sources and merged into a single dataset.
	Google Map Maker (2017)	Point vector (updated until 2014)
Class 4: Access to adequate Sanitation – UN Habitat requirement
Pit latrines (1)	OpenStreetMap (2017)	Point vector (updated until 2014)	Slums have a large number of pit latrines.	The XY location of pit latrines was extracted from both data sources and merged into a single dataset.
	Google Map Maker (2017)	Point vector (updated until 2014)
Class 5: Security of tenure – UN Habitat requirement
Online real estate activity (1)	Online real estate agencies in Kenya and Africa (e.g. Property24 – www.property24.co.ke)	XY location and attribute characteristics of online real estate activity – rental, sale or purchase of house, commercial property, building or land – (2010–2014)	Slums, due to their typical illegal occupancy of the land in which they are built, have little to no online real estate activity.	The XY locations of real estate transaction locations were collected from various online websites and merged into a single dataset.
Class 6: Auxiliary Candidate Indicators
Quality of life – NDVI (1)	Landsat 8 – Image bands 4 and 5 (Google Earth Engine 2017)	30 m spatial resolution (2014)	Slums have low amounts of vegetation compared to other settlement types.	This dataset was created by generating an NDVI layer for each set of Landsat imagery collected for the year 2014. The median value for each pixel across this stack of NDVI images was then extracted to create a greenest pixel NDVI dataset. Greenest pixels values were rescaled to values between 0 and 1 to have only positive instances of this indicator.
Roads – poor quality (1)	OpenStreetMap (2017)	Polyline vector (2014)	Slums have a high quantity of poor road types.	Poor quality roads (i.e. unpaved, unsurfaced, track, path, footpath, gravel) were extracted from each data source and merged into a single dataset.
	Google Map Maker (2017)	Polyline vector (2014)
Roads – dead ends (1)	OpenStreetMap (2017)	Polyline vector (2014)	Slums have a high quantity of dead-end streets.	Road data for both datasets were merged into a single layer. Nodes in this street layer with only one connected edge were considered dead-end nodes. These nodes were extracted.
	Google Map Maker (2017)	Polyline vector (2014)
Roads – average curvature (1)	OpenStreetMap (2017)	Polyline vector (2014)	Roads in slums have high average curvature values (due to poor urban planning practices) indicating more irregular roads.	An algorithm was written to examine the vector angle between every 3 nodes in the road layer. For example, if a segment of road contained 5 nodes (n1, n2, n3, n4 and n5), then angles were calculated as follows: n1, n2, n3; n2, n3, n4; n3, n4, n5. The average of these angles was then computed for each road segment, where a road segment was represented as the segment between two intersection nodes, one intersection and one dead-end node, or two end nodes.
	Google Map Maker (2017)	Polyline vector (2014)
Roads – total length per unit area (1)	OpenStreetMap (2017)	Polyline vector (2014)	Slums have lower road densities per unit area when compared to other settlement types.	The total length of roads from the two data sources was merged and computed for each 1 km by 1 km grid cell.
	Google Map Maker (2017)	Polyline vector (2014)
Road – roads per unit population (1)	Road – OpenStreetMap (2017)	Polyline vector (2014)	Slums have lower amounts of roads per unit population.	The total length of roads per 1 km  × 1 km grid cell were divided by the population density contained within that grid cell.
	Road – Google Map Maker (2017)	Polyline vector (2014)
	Population – LandScan (ORNL 2017)	1 km by 1 km spatial resolution (2014)
Services – private primary schools (1)	Kenya Open Data (KOD 2018)	Point vector (updated until 2014)	A large number of private primary schools are located in and immediately around slums.	XY locations of private primary schools were extracted from the data source.
Services – private health care facilities (1)	Kenya Open Data (KOD 2018)	Point vector (updated until 2014)	A large number of private health care facilities are located in and immediately around slums.	XY locations of private health care facilities were extracted from the data source
Services – places of worship (1)	OpenStreetMap (2017)	Point vector (updated until 2014)	A large number of places of worship are located in and immediately around slums.	Places of worship (e.g. churches, masjids and temples) were extracted from each data source and combined.
	Google Map Maker (2017)	Point vector (updated until 2014)
	Kenya Open Data (KOD 2018)	Point vector (updated until 2014)
Ambient Geospatial Information – Flickr (1)	Flickr API (Flickr 2017)	Location of Flickr images containing the keywords ‘slum’ or ‘informal settlement’ in their tags (2010–2014).	A large number of geolocated Flickr points containing the keywords ‘slum’ or ‘informal settlement’ in their tags are clustered in and around slums.	A script was written to query the Flickr API using search terms ‘slum’ or ‘informal settlement’ The XY locations of features resulting from this search query were then merged into a single dataset.
Local news media – text mining (1)	Local online news media sites in Kenya (e.g. Daily Nation – https://www.nation.co.ke)	Online news articles discussing slums in Kenya (2010–2014).	A large number of news media articles on slum settlements are based on reported activity within and around slums.	Articles from Google News were curated using the search terms ‘slum’ or ‘informal settlement’. A script was written to then extract the sentence in each script containing these search terms. The overall assumptions were that the name of the slum would be likely to occur within this sentence of words. Sentences were then individually tokenized, their nouns and proper nouns extracted, and these further geocoded using both OSM and Google Geocoding services. Geocoding was restricted to the specific study areas to reduce the instance of false positives. The locations of features resulting from successful geocoded names were then combined into a single dataset.

Figure 3. Distribution of positive classified cases for slums for (a) logistic regression, (b) discriminant analysis and (c) the See5 decision tree.

Table 2. Distribution of slums cases.

Study area	Slum cases (% of total)	Non-slum cases (% of total)	Total number of cases
Nairobi	3.40	96.60	1,205,728
Mombasa	8.48	91.52	1,048,407
Kisumu	1.50	98.50	717,545

Table 3. Overall classification accuracies.

Model	Nairobi		Mombasa		Kisumu
	Precision	Recall	Precision	Recall	Precision	Recall
LR	0.01	0.10	0.29	0.57	0.45	0.72
DA	0.66	0.22	0.79	0.35	0.71	0.36
See5	0.90	0.98	0.93	0.95	0.93	0.94

Table 4. Indicators for first four levels for See5 tree for each study site.

Candidate indicators	Nairobi					Mombasa					Kisumu					All sites
	Tree level				Total	Tree level				Total	Tree level				Total
	1	2	3	4		1	2	3	4		1	2	3	4		Total
Socio-economic
Average road curvature	0	0	0	0	0	0	0	1	2	3	1	0	1	0	2	5
Births	1	0	0	0	1	0	0	0	0	0	0	1	0	0	1	2
Dead ends	0	0	0	1	1	0	0	0	0	0	0	0	0	4	4	5
Population	0	0	2	0	2	0	0	0	0	0	0	0	0	0	0	2
Real estate activity	0	0	0	0	0	0	0	1	1	2	0	0	2	0	2	4
GeoSocial media	0	0	1	0	1	0	0	0	1	1	0	0	1	0	1	3
Services	0	1	0	2	3	1	0	0	1	2	0	1	0	0	1	6
Total per level per site	1	1	3	3	8	1	0	2	5	8	1	2	4	4	11	27
Satellite-derived indicators
Greennest	0	0	0	1	1	0	0	1	0	1	0	0	0	0	0	2
Entropy 13 × 13	0	0	0	0	0	0	0	0	1	1	0	0	0	0	0	1
Second moment 9 × 9	0	0	0	1	1	0	0	0	0	0	0	0	0	0	0	1
Second moment 13 × 13	0	0	0	1	1	0	0	0	0	0	0	0	0	0	0	1
Homogeniety 13 × 13	0	0	0	0	0	0	1	0	0	1	0	0	0	0	0	1
Mean3 × 3	0	0	1	1	2	0	0	0	0	0	0	0	0	0	0	2
Mean 5 × 5	0	0	0	0	0	0	0	0	0	0	0	0	0	1	1	1
Mean 7 × 7	0	1	0	0	1	0	0	0	0	0	0	0	0	0	0	1
Mean 9 × 9	0	0	0	0	0	0	1	0	0	1	0	0	0	0	0	1
Mean 11 × 11	0	0	0	0	0	0	0	0	1	1	0	0	0	0	0	1
Mean 13 × 13	0	0	0	0	0	0	0	0	0	0	0	0	0	1	1	1
Total per level per site	0	1	1	4	6	0	2	1	2	5	0	0	0	2	2	13
All indicators	1	2	4	7	14	1	2	3	7	13	1	2	4	6	13	40

Figure 4. Urban and slum areas in Nairobi (False composite image created by stacking image bands 7, 6 and 4 from the Landsat 8 satellite. Pink to white tones highlight urban areas).

Table 5. Percent change in accuracy using the remote sensing data as the baseline.

RS + 1	RS + 2	RS + 3	All
6.55	6.81	6.85	6.52
1.55	2.38	2.51	2.49
1.62	2.05	2.087	2.13

Note: Number in the top row represent the addition of the top 1, 2 and 3 socio-economic indicators to the remote sensing data. Where ‘RS’ represents the remote sensing indicators and ‘+1’, ‘+2’ and ‘+3’ represents the addition of the single, top 2 and top 3 socio-economic indicators, respectively.

Figure 5. See5 tree up to 4th level for (a) Nairobi, (b) Mombasa and (c) Kisumu.

Figure 6. ROC curves for See5 trees built with the remote sensing indicators, the progressive addition of the most significant socio-economic indicators to the remote sensing indicators, and using all indicators for all three study sites. The inset image in each figure shows the full distribution of the data.

Table 6. Data mining models applied to other study areas.

Sites	Nairobi		Mombasa		Kisumu
	Precision	Recall	Precision	Recall	Precision	Recall
Logistic regression
Nairobi	NA	NA	0.00	0.07	0.00	0.00
Mombasa	0.01	0.19	NA	NA	0.00	0.01
Kisumu	0.03	0.18	0.22	0.29	NA	NA
Discriminant analysis
Nairobi	NA	NA	0.34	0.30	0.07	0.06
Mombasa	0.8	0.14	NA	NA	0.46	0.30
Kisumu	0.17	0.18	0.73	0.22	NA	NA
See5
Nairobi	NA	NA	0.25	0.34	0.03	0.42
Mombasa	0.37	0.12	NA	NA	0.08	0.16
Kisumu	0.46	0.14	0.46	0.14	NA	NA

Note: NA: not applicable.

OpenStreetMap. 2017. Accessed January 12, 2017. http://www.openstreetmap.org.

 Google Scholar

Google Map Maker. 2017. Accessed February 20, 2017. https://services.google.com/fb/forms/mapmakerdatadownload/.

 Google Scholar

Google Earth Engine. 2017. Accessed February 20, 2017. https://earthengine.google.com.

 Google Scholar

ORNL. 2017. Landscan. Accessed February 12, 2017. http://web.ornl.gov/sci/landscan/.

 Google Scholar

WorldPop. 2017. “WorldPop Project.” Accessed May 25, 2017. http://www.worldpop.org.uk.

 Google Scholar

KOD. 2018. “Kenya OpenData” Accessed May 21, 2018. https://opendata.go.ke.

 Google Scholar

Flickr. 2017. Accessed February 15, 2017. https://www.flickr.com.

 Google Scholar