Exploration of Terrace or Linked Houses Detection from Satellite Images Using Deep Learning

This paper presented a study about Malaysian terrace of linked houses detection from satellite images using the YOLOv8 segmentation model. The study covers various crucialaspects including dataset preparation, annotation, data augmentation, data partitioning, network architecture selection,training, and performance metrics. The dataset was first built by collecting 100 satellite images from Google Earth, focusing on Malaysian terrace or linked houses. Then, annotation was performed using polygons to provide accurate ground truth data for training. Different data augmentation techniques wereexplored and applied to increase the diversity and robustness of the training data. The dataset was then divided into training, validation, and testing sets for unbiased evaluation. The YOLOv8 model was selected for this study due to its simplicity, efficiency, and consistent object localization capabilities. The performance of the trained model was evaluated using metrics such as Intersection over Union (IoU), confusion matrix, and Mean Average Precision (mAP). The results showed that increasing the dataset size and epochs improved the mAP values. This paper also discussed the limitations and potential future research directions in this field. Overall, the trained model contributes to the accurate detection of buildings from satellite images with highest mAP values of 0.62.