House Boundary Detection in South Asia

Project Title

House Boundary Detection in South Asia

Project Area of Specialization Artificial IntelligenceProject Summary

This project will be using the Multi-temporal Urban Development SpaceNet (MUDS, also known as SpaceNet 7) dataset and the VGG 16 network to predict housing boundary using a custom a loss function especially catered for landscapes of South Asia, in particular Pakistan. We aim to exploit the fact the houses in cities such as Karachi and Lahore have a fixed land area, even if there is no visible boundary to differentiate between the two. We will use this to create a loss function that predicts house boundaries on these images. 

Project Objectives

This project aims to help identify houses in the satellite images of Pakistani cities. This will allow us to track change in landscape, predict population densities, and also help create data and statistic driven policies. Another important objective that this will allow us to achieve is to compare tax filings with the land area of each house and can automate the process of identification of tax evaders. 

Project Implementation Method

We will be using the spacenet 7 dataset primarily to train our model, using different researched loss functions. This requires access to the Solaris framework provided by AWS (Amazon Web Services). We will be comparing the results with the baseline model and documenting the F1 scores and accuracy of our model. The project will be using pytorch, keras, openCV and other python libraries and will also require access to the GPU and TPU provided by google colab. 

Benefits of the Project

The project allows for predictive analysis of population growth and city expansion. It allows for resource management and enables the policy makers to track changes in infrastructure in a city or area. It will even allow to automate the process of evaluating the tax fillings with land area. 

The project is also beneficial as we aim to make the model perform high accuracy boundary detection on low resolution satellite images, the use of existing infrastructure to enable complex analysis can potentially allow the cost of this analysis to be much lower.

Technical Details of Final Deliverable

The final model will be a Solaris enabled VGG 16 Unet architecture with a custom loss function to detect house boundaries of South Asia, particularly Pakistan. The model will be trained on the spacenet 7 dataset and assessed on the satellite images of Lahore. 

We also aim to publish our findings in a research paper, especially our approach to the custom loss function which is the most important and critical objective of our project.

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

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	57296
Google Colab Pro	Equipment	2	20000	40000
Google One Standard Plan	Equipment	2	5948	11896
AWS root account	Equipment	2	200	400
Overheads	Miscellaneous	5	1000	5000