House Boundary via satellite imagery

Project Title

House Boundary via satellite imagery

Project Area of Specialization Artificial IntelligenceProject Summary

High taxes, traffic congestion, accidents, pollution, and anxiety are all urban operational and social vulnerabilities that can be mitigated with effective urban development and management planning. Estimation of building footprints is essential for urban studies and application. Identifying changes in and use of urban centres can prove useful in improving tax collection, estimating number of households, identification of unauthorised construction, and planning infrastructure development such as schools, hospitals and mass transit network.

Currently available methods to extract building footprints are very expensive, time consuming, and hampered by human and instrumental error, due to intensive-manual intervention. Furthermore, these methods are only applicable to well-structured urban housing in developed countries, and do not accommodate well to challenges associated with data from developing countries, such as Pakistan. Therefore, urban development planning developments have to spend a large amount of resources in conducting manual surveys, every three years, to maintain a database for building footprints. Goal of this project is to automate this process.

While it is easy to distinguish developed land with unutilized spaces, the more important issue is what is the spatial extent (over land) of each building, or property, and how is it changing over time? In this project, we aim to provide a solution to this issue, which is related to the spatial extent of building, which we can refer to as its footprint. Unlike most modern solutions, which are highly data-driven, which don’t translate well to Pakistan, we aim to combine contextual information such as access to the road, and the available data. The aim is to obtain a generalised, and interpretable solution that can be applicable in local scenarios, which are usually marred by ill-planned development that suffers from congestion, and irregular growth.

Project Objectives

With advancement in modern machine learning, many methods have been proposed for automated survey, via satellite imagery, and extracting building footprints. These methods are purely data-driven i.e. provide a solution that does not incorporate the known mathematical models that exist in literature to estimate building footprints, therefore these methods lack generalisation. This project aims to combine both the known mathematical frameworks with data-driven machine learning to propose a solution for estimation of building attributes, which includes its footprint, and changes in that over time. We will be utilising both multi-spectrum and multi-view imagery to improve the robustness and generalisation of our approach so it can be made available in local scenarios. The objectives of this project are as follows:

1. The first objective is to build the database for the aforementioned problem, based on multi-modality and multi-view cohorts gathered from different resources and make it available for concerned authorities.
2. The developed database from the first component would be utilised to develop automated algorithms for building attribute estimation.
3. Extract building dimensions and estimate the covered area of the building from the satellite imagery, while incorporating surrounding contexts, which includes access to roads, neighbourhoods, and uniformity constraints.

Project Implementation Method

We aim to utilise our existing knowledge and algorithms, and propose state-of-the-art efficient and generalised modules to estimate house boundaries. The method is as follows:

1. In the first stage, data gathering will be done. Here, we will be digitising existing real-estate maps, and data will be incorporated to obtain ground truth which gives us information including: PHASE, SECTOR, TYPE (Residential, Commercial etc), HOUSE_NO, Shape_Length, Shape_Area, Latitude, Longitude, and Geometry
2. After data has been gathered, we will generate our own annotations using Google Earth, and QGIS. Additionally, data from a physical survey conducted by colleagues at the Economics Department will also be utilised.
3. Houses in a housing block exhibit some common features, such as they all have access to roads, follow some grid-like placement, and mostly similar in plot size. In this step, we will incorporate this information into our learning framework. This will be done by designing a cost function that will have a weighted sum of five components, once for each size (s), position (p), variance in size (?), uniform layout (u), and access to road (r). The function can be defined as: L = ?s + ?p + ?? + ?u + ?r
4. To enable this loss function, we will develop a coarse to fine strategy that first extracts a housing block, and then predicts boundaries for each house in the block. 
5.  For each bounding box, we will compute all five terms of the loss function. The “s” and “p” will be computed from the ground truth, “?” will be computed using the sizes of all the predicted boxes, “u” will be computed by estimating the gaps between the house and differences in their position, and “r” will be obtained using the housing block boundary.
6. The above designed loss function can be implemented through existing object detection frameworks that only work on the size and position. These include YOLO, SSD, and Efficient Net. Then we build a Convolution neural network with novel loss function that incorporates the mathematical models along with structural constraints to improve the accuracy and make the solution applicable in challenging local circumstances. We will utilise well-known baseline models like YOLO to test our customised loss function. This phase will result in an efficient solution for house boundary extraction.

Benefits of the Project

The project will result in development of a new machine learning strategy, which is more interpretable and generalizable due to its association with mathematical foundations. The developed technology will automatically generate land-used/land-covered, and will help in identifying regions and locations where land use has changed over time (for example, from no building to building, or multiple buildings existing where there was only one).

This will help improve revenue collection, by improving accuracy in identifying and classifying properties leading to higher property tax revenues, as well as transparency for all stakeholders by providing them with tools, with which they can identify property tax revenue discrepancies, which leads to improved tax collection.

This project can also be extended to the concept of smart cities, where a detailed digital footprint of them exists, allowing for maintaining and managing them via data-driven decision making.

Technical Details of Final Deliverable

The final deliverable would be combining the entire algorithm with a software front end, for usability testing. This will essentially take satellite images, taken from Google Maps, and then estimate house boundaries in the input image, listing the amount of households estimated, and allowing users to view any particular house boundary. The software will also classify the house boundaries into categories such as residential, commercial etc.  We can provide the software with multiple images of the same area, but taken over a period of time, let’s say one year, to see if there were any changes in land covered, and house boundaries, which would then be used by tax collection departments to update the property tax of that area. The software will also provide an estimate of property tax for a given house boundary.

Final Deliverable of the Project Software SystemCore Industry ITOther IndustriesCore Technology Artificial Intelligence(AI)Other TechnologiesSustainable Development Goals Decent Work and Economic Growth, Sustainable Cities and CommunitiesRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	70000
Data storage with	Equipment	1	20000	20000
GPU Machine without GPU	Equipment	1	50000	50000