Land Cover Extraction from Satellite Imagery using U-Net Model

Satellite imagery offers extensive information that can be used for a variety of societal applications, from the number of buildings in a metropolis to the land cover types of a specific area. However, extracting such information is a difficult task. In the past, images were scanned manually by domain experts to extract features like buildings, infrastructure, and vegetation, which would take weeks or even months depending on the image size. To address this challenge, we trained our model using U-Net architecture, a deep learning model, because of its advantages in multiscale feature learning and faster processing. Data augmentation is done on the dataset in order to enhance the amount of data and prevent overfitting artificially. The trained model achieved an F1 score of 92% and an impressive precision score of 95%. This automated method can be scaled to large sets of data, reducing the need for human labeling.