Abstract:
his research thesis enhances the state-of-the-art building detection in very high resolution aerial images by fine-tuning pre-trained state-of-the-art deep models on three
diverse benchmarking datasets: the Massachusetts Buildings Dataset, WHU Buildings
Dataset, and INRIA Aerial Image Labeling Dataset. Through the use of Dual-stream
Asymmetric Fusion Networks, the Dice Loss, IoU Score, Precision, Recall, Accuracy,
and F1 Score were enhanced. Our approach was to tune the existing model architecture
using different datasets to enhance generalization capability and accuracy. Performance
after training and validation consistently showed improvements, indicating that the model
was effectively learning and enhancing the building segmentation accuracy. Visual
results indicated the model's robustness to different urban environments, where the
ground truth matched the predicted masks. These advancements can have high
importance for urban planning, disaster management, and GIS, where precise building
detection becomes an important criterion. This work emphasizes the importance of using
advanced model architectures in conjunction with comprehensive data integration to
enhance the building detection system. The improvements achieved prove the efficiency
of our approach in real-world applications and greatly contribute to remote sensing and
building detection. Our research entails practical solutions and insights, paving the path
for future improvements in detecting buildings in aerial images accurately and efficiently.
