Urban Scan

Over the past few years, I have noticed many construction projects in my city and nearby areas. This hackathon made me think about how these projects affected the environment and if there was any way to find their impacts beforehand. This led me to create Urban Scan, an application that uses satellite imagery to determine the environments present in the area and how the different ecosystems are affected. It uses an image classification model to analyze the image and label the different environments detected. One challenge that I ran into was finding a good model. In order to accurately detect the shape and size of various ecosystems, I had to determine the environment type of each pixel based on the color instead of classifying a cluster as a single environment. In the beginning, I tried to see if I could have preset conditions to classify a pixel as a color based on the RGB values, but I found this to be inaccurate as I could not take into consideration the thousands of possible values. I tried using an image classification mode next, but as pre-trained models are only able to detect common objects, I had to train my own model. After using hundreds of images to train the model and days spent on adjusting the backbone, I was able to create an accurate image classification system. I found that a simpler backbone was required as looking too much for features within the image resulted in more time taken to process the image as well as a decrease in accuracy. I plan to improve this application by taking into account animal populations in various parts of the image to find what other animals might live in the area. This will help me better determine the effect urbanization could have on plant and animal life.