For our final year project, our desire was to automate an existing manual system using state-of-art Machine Learning models and provide results with high accuracy so that it is productionable. One of the main problems in a developing metropolitan city like Mumbai, India was careful planning and growth of poorly developed areas. Currently, Municipal Corporation staff would manually collect data by physically visiting the places and would keep a record in GIS software. We decided to work on the problem statement to identify slum areas in real-time using satellite images of any given region.
Our first aim was to research similar projects, and how they performed object identification. We had to identify and select a model which would be best suited for such image data. We researched a number of models and did a comparative study on many of these models. (https://www.ijntr.org/download_data/IJNTR05020014.pdf ) After a few trials, we decided to go with the Mask R-CNN model since it was the most robust and accurate model for our application.
Next, we had to create the dataset of images for training the model. For training, we required images having annotations of slum areas so that our model learns to distinguish and identify them. To get the most accurate information on slum areas, I visited the Municipal office to understand how they actually store the data. We then got access to the Georeferenced Slum Clusters in geojson format. To understand how to make use of this data, I attended a 10-day Workshop on GeoSpatial database. Instead of manually annotating the images, I investigated how we can use this cluster information available to get the required dataset. After much deliberation, I finally developed an automated script to annotate the images using GIS Software and to download them on the hard disk for training purposes.
We trained our model using Nvidia GeForce GTX 650 Cuda configuration using 8GB RAM and 1TB HardDisk Later, we created an end-to-end product that would allow the user to select any desired area of interest on Google Maps and obtain the identified slum with confidence percentage.