Using Satellite Imagery For Good: Detecting Communities In Desert and Mapping Vaccination Activities
Our algorithm detects areas that are clustered together (to represent housing-clusters) and polygon fitting is performed so that we can create manageable representations (you can visualize our results on the interactive map below and technical details are discussed in the last section). Finally we map vaccination and public school location data over these detected regions. For vaccination data we calculate both count of vaccination activity per quarter per region, as well as coverage to understand how geographically diverse activities were performed in those region. Our analysis shows that although not all the detected regions appear to be covered, large number of them are. Some get vaccination activity in all the quarters, others are missed in one or two of the quarters. Large number of detected regions had at-least one public school as per government record. Certain detected regions were found to be getting very few or none of vaccination activity, their location and cluster indicates to us that there could be anomaly in the data, we hope to verify and correct that in our future version. We are in process of conducting such survey for whole Punjab, on much higher resolution satellite imagery.
Correlating vaccination activity with detected communities
As per Government's vaccination programme each vaccinator maintains a record that contains age of children with there pending vaccines. Following this record vaccinators visit their assigned vicinities to complete a vaccine's round according to its specificity. During each activity a histogram is plotted showing number of vaccinators who accessed a detected community. These numbers justify the fact that regions with larger areas were accessed the most.
Just like vaccination activity information, government schools data was correlated with the detected regions to understand accessibility to education. It was discovered that a large number of the detected regions did have the government school in them. Red areas show polygons with no schools, where as the shades of green represent increasing number of schools with light green being the least(one or two) and dark green with most(more then 10) schools.
We have used village finders satellite imagery data in order to train our network. Image tiles are divided in small chips of size 128 and further resized to 64. Data augmentation methods were used before and during training. The probability map achieved after last convolution layer is bilinearly interpolated to achieve input size. A threshold of 0.5 is applied to generate a binary image. Polygons are fitted on these binaries resulting in detected communities. The deep learning program is written in python using Chainer framework.
A shortcoming is that our algorithm cannot predict height or even type of the building. Therefore some of the building detected are themselves private schools run. Furthermore some built structures are missed that is because they seem to have merged into the landscape when viewed at certain zoom level.
Satellite imagery offers a vast research area. To take full advantage of the diversity these images contain, we plan to study the variety of bands that hyperspectral imagery presents. Detecting built structures successfully has inaugurated problems like house counting, tracking urbanization and detecting damaged structures. We intend to deeply analyze the detected segments in order to locate hospitals, schools and parks near each community.
We are grateful to Punjab Information Technology Board (PITB) especially Ms. Maria Badar for sharing vaccination activity data as well as public schools data. We would also like to thank Muhammad Umer Ramzan, ITU's management and their website creators for helping us build this page.