Building Detection Using Convolutional Neural Network (CNN) U-Net Architecture on Satellite Imagery

This research aims to detect buildings in satellite images using the U-Net architecture CNN model. The dataset is a satellite image of the Araya area, Malang City extracted through Google Earth Engine with the Python programming language, consisting of 4 band channels: Red, Green, Blue, and NIR. The data is processed into 2 types of sizes, namely 128 × 128 pixels and 64 × 64 pixels, then divided into train, validation, and test data. Training uses Keras Tensorflow, and testing is done in three stages: model performance testing with layer addition and parameter changes, building detection testing, and comparison of prediction results with manual measurements. The best model was obtained by a model with a training dataset of 128 × 128 pixels that achieved 93% accuracy with a prediction time of 2 seconds.