skip to Main Content

Monitoring Forced Evictions in Phnom Penh, Cambodia Using High Spatial Resolution Satellite Imagery

By: Zachary J. Bortolot, Department of Integrated Science and Technology, James Madison University, GISCorps Volunteer

Kathryn R. Striffolino, Amnesty International USA

Nora Lindstrom, Sahmakum Teang Tnaut


Since 2000, over 100,000 families have been forcibly evicted from their homes in Phnom Penh, Cambodia (STT, 2009).  After being evicted, the families are typically left homeless or relocated to homes outside central Phnom Penh which lack basic amenities and where opportunities for work and access to education and healthcare are limited (STT, 2009; Amnesty International, 2010).  This is in direct violation of international human rights law, including the right to adequate housing enshrined in the International Covenant on Economic, Social and Cultural rights to which Cambodia is a party.

In order to raise awareness of the evictions, a local non-governmental organization, Sahmakum Teang Tnaut (STT), has been collecting data on the locations where the evictions have taken place and the number of families that have been evicted since 1997.  These data have been collected by surveying trusted community contacts in the affected communities (STT, 2009) and by sending STT staff members to eviction sites with a GPS device to record the locations of the eviction sites.

Although these data points are collected carefully and are considered to be very reliable by STT, an independent validation of the data based on satellite imagery was desired in order to better make the case to the international community that the estimates are valid, that local and international action should be taken, as well as attempt to provide irrefutable evidence of the destruction of homes for use in redress and accountability efforts.  Unfortunately the scale of the project meant that the traditional approach of manually counting the number of demolished buildings in the eviction areas was not feasible.  Therefore earlier this year STT and Amnesty International USA asked GISCorps for assistance.  GISCorps agreed, and the result was that the three authors of the paper were able to join together by e-mail and in person at Amnesty International USA’s office in Washington, DC to develop and apply a methodology for independently estimating the number of families evicted from eviction areas in Phnom Penh between 2006 and 2012.  This report presents the method that was developed for doing this and the results which were obtained.

In order to estimate the number of evicted families, two high spatial resolution satellite images were acquired which covered the four central  khan (Figure 1).  The first was a Quickbird-2 image acquired on July 5, 2005, and the second was a WorldView-2 image acquired on July 12, 2012.   These images were selected because they were virtually cloud free and they were collected at near anniversary dates which reduced seasonal variations between the images.

Once the images were acquired it was discovered that the alignment was poor (Figure 2).  To correct this, the 2005 image was orthorectified using PCI Geomatica 2013.  The 2012 imagery was used as the reference image and an Aster Global DEM[1]  was used for elevation.  For the orthorectification, 50 ground control points were collected, and the resulting RMSE was 1.14 m.  The resolution for all bands in the 2005 imagery was set to 50 cm.

After orthorectification was performed, GPS point data showing the locations of the 24 communities that had been evicted between 2006 and 2012 were loaded, and heads up digitizing was used to create polygons reflecting the boundaries of the different eviction areas based on the 2005 and 2012 images (Figure 3).  The communities evicted in 2005 were not included because the imagery was collected in July, 2005, and in the communities evicted in 2005 it is likely that some demolition may have begun prior to the image collection date.  This would have led to inaccurate results.  Due to uncertainty regarding the boundaries between adjacent eviction areas, a number of eviction areas were combined together.  The net result was 13 areas in which evictions had occurred.  The boundaries were evaluated by staff at STT who had extensive knowledge of the eviction areas, and they said that the boundaries appeared to be accurate.

Figure 1.  The location of the study area.

 Figure 2. A figure showing the misalignment present between the 2005 and 2012 images.  The misalignment was fixed through orthorectification.  DigitalGlobe Imagery Provided by eMap International.

Figure 3. A sample of the boundaries surrounding an eviction area.  The image on the left is from 2005, the one on the right is from 2012.  DigitalGlobe Imagery Provided by eMap International.

The next step was to perform a classification to identify pixels within the demolition areas that contained buildings in 2005.  To do this, a two stage ISODATA unsupervised classification was performed in PCI Geomatica 2013 using texture metrics calculated from the 2005 imagery (see Table 1).  For the first stage, 250 classes were generated within the boundaries of the eviction areas and each class was identified through visual interpretation as being either being part of a building, not part of a building, or a mixture of building and non-building areas (Figure 4).  For the second pass, just the pixels which were identified as belonging to the mixed class were classified into 250 classes and each class was identified as being part of a building or not part of a building.  The net result was an image showing the locations of buildings in the 2005 image (Figure 5).

After the classification was performed, the boundaries of buildings that remained in 2012 (i.e., that were not demolished) were digitized manually based on comparing the 2005 and 2012 images (Figure 6).  These results were combined with the classification results to produce an image showing just the pixels containing buildings that had been demolished between 2005 and 2012.

To assess the accuracy of the image containing the buildings that were demolished between 2005 and 2012, 229 pixels within the eviction areas were randomly selected and a confusion matrix was constructed based on visual interpretation (Table 2).  The overall accuracy of the classification was found to be 84.7%.  The method described by Wynne et al. (2000) was then used to find the bias adjusted demolished building area in each eviction area and the associated 95% confidence limit.

Next, 259 pixels that were classified as being part of demolished buildings were randomly selected.  This number was derived by sampling 0.02% of the pixels classified as being part of a demolished building, and then adding additional random points as needed to ensure that each eviction area contained at least five selected pixels.  The boundary of the nearest demolished building was then digitized based on the 2005 image.  In order to help with interpretation Bing Maps imagery was consulted in some cases.

Once the buildings had been digitized the weighted mean and weighted variance of the building areas in each eviction area were calculated.  The weights were set based on the following equation:


Where W is the weight associated with a building.

The weighting was necessary due to the way buildings were selected since larger buildings had a higher probability of being selected than smaller buildings (e.g., the randomly selected pixel was four times more likely to land in a 100m2 building than in a 25m2 one).  The mean building size and 95% confidences limits of the building sizes in each eviction area were then found.

In each eviction area, the area classified as containing demolished buildings was divided by the mean area per building in that eviction area.   This gave an estimate of the number of demolished buildings.  To calculate the confidence limits associated with the number of buildings in each eviction area, the error propagation method for quotients of two variables having independent random errors was used (Taylor, 1997).

To estimate the number of families that were evicted, representatives from STT visited five communities that have not been evicted but which are considered to be representative of the ones that have.  In these communities they visited all buildings (427) to find out the number of families living in each one.   These data allowed the mean number of families per building and the associated 95% confidence limit to be calculated.

The estimated number of families displaced and the 95% confidence limit was calculated by multiplying the estimated number of demolished buildings by the mean number of families per building.  The error propagation method for the product of variables having independent random errors was then used to combine the confidence limits (Taylor, 1997).

Table 1. Values used as input to the classification.  All inputs were normalized prior to classification by rescaling them to a range of 0 to 255.

Table 2. The confusion matrix based on 229 randomly selected points.

Figure 4. The process of unsupervised classification.  In the first stage of the classification each class was identified as being either being part of a building, not being a part of a building, or being a mixture of the two.  The pixels classified as belonging to the mixed class were then input into a second unsupervised classification (stage 2).  DigitalGlobe Imagery Provided by eMap International.

Figure 5. A sample of the classified image.  The green pixels indicate the locations of buildings in the 2005 image.  DigitalGlobe Imagery Provided by eMap International.

Figure 6. The outlines of buildings in eviction areas that were present in both the 2005 and 2012 images were manually digitized and removed from the classified image (Figure 5).  This resulted in an image showing the locations of buildings that had been demolished between 2005 and 2012.  DigitalGlobe Imagery Provided by eMap International.

The total number of demolished buildings and the total number of families displaced in all areas combined was estimated by summing the estimates obtained from the individual eviction areas.  The error propagation method for the sum of variables having independent random errors (Taylor, 1997) was used to calculate a combined confidence limit.

Results and Discussion

The results of the analysis are shown in Table 3 and show that for 62% of the individual eviction areas, the survey-based estimates of the number of displaced families are within the 95% confidence limits associated with the satellite imagery-based estimates.  For all eviction areas combined, the estimate of the total number of families displaced based on surveys differs from the satellite-based estimate by 1%, which is within the 95% confidence limit associated with the satellite-based estimate.  There are a number of possible reasons for the discrepancies between the satellite and ground based estimates.  Possible reasons include the following:

  1. It was assumed that people were evicted from a settlement area when a building was demolished.  However, in some eviction areas people were relocated into newly constructed buildings.
  2. The estimated number of people per building was based on data from just five areas that had not been demolished.  It is therefore likely that some of the eviction areas may have had substantially higher or lower numbers of families per building.
  3. Many of the image processing steps relied on human interpretation and therefore may contain errors.


This study used multitemporal satellite imagery to estimate the number of families that had been displaced from areas in which forced evictions had taken place in order to independently validate data collected on the ground using surveys.  The results show that the majority (62%) of the survey based estimates for the individual eviction areas are within the 95% confidence intervals associated with the satellite based estimates, and the total number of evicted families obtained using satellite data differs from the estimate obtain using surveys by only 1%.  This strongly suggests that the estimates of the number of displaced families obtained using surveys are reliable.  In the near future the authors are planning to use a similar methodology to estimate the number of families living in areas which have received eviction notices but which have not yet been evicted.

It is hoped that this independent confirmation of the ground survey data will be useful in making the case that mass evictions are taking place in Phnom Penh, and that this will ultimately lead to the cessation of forced evictions and to an improvement in the lives of Phnom Penh’s inhabitants.  The authors are very grateful to GISCorps for its help in making this study possible, and overall it was an excellent experience for all three authors.

Table 3. Preliminary results of the analysis which are subject to change.  The locations and names of the different eviction areas will be released once the results have been finalized.


Amnesty International , 2012.  Cambodia: Ongoing Serious Human Rights Violations Must be Addressed.  Available online at

Sahmakum Teang Tnaut, 2009.  The 8 Khan Survey: Urban Poor Settlements in Phnom Penh.  Available online at

Taylor, J. R., 1997.  An Introduction to Error Analysis.  Second edition.  Sausalito, CA: University Science Books.  327pp.

Wynne, R. H., Oderwald, R. G., Reams, G. A., Scrivani, J. A., 2000.  Optical remote sensing for forest area estimation.  Journal of Forestry 98(5): 31-36.

[1] ASTER GDEM is a product of METI and NASA

Back To Top