Assessing the cross-region knowledge transfer capability of selected deep learning building vectorization methods in...

The paper analyzes the cross-region knowledge transfer characteristics of selected state-of-the art deep learning-based building vectorization methods in the context of their architecture and choice of training dataset. The goal of this work is to assess whether certain combinations of deep neural network models and training datasets deliver a higher level of cross-region knowledge transfer when applied to processing of Earth Observation images sourced from geographic locations that lack high-quality training data. The study investigates representative state-of-the-art neural network architectures, including CNN (represented by HiSup), Transformer (represented by Pix2Poly), GNN (represented by PolyCity) and general-purpose models (represented by SAMPolyBuild), using the SpaceNet AOI 2 Las Vegas, SpaceNet AOI 5 Khartoum, CrowdAI, and WHU-Mix datasets for training and a custom dataset representing Dar es Salaam, Tanzania, for testing. The results indicate that a network's cross-region knowledge transfer capabilities are primarily affected by its segmentation and vertex selection methods, and thus networks that exhibit similar performance on training datasets can show significantly different cross-region knowledge transfer capabilities. Moreover, while the cross-region knowledge transfer performance of tested building vectorization networks is strongly influenced by the size and diversity of training data, training on high-quality datasets from different regions usually produces significantly better results than using a lower-quality dataset from a similar area. Among the evaluated methods, the best overall performance was demonstrated by SAMPolyBuild trained on the WHU-Mix dataset, which achieved Average Precision (AP), Average Recall (AR), and IoU scores of 31.6%, 37.9% and 68.9%, respectively, versus AP of 60%, AR of 65% and IoU of 84% when evaluated on the test subset of WHU-Mix. In individual metrics, Pix2Poly achieved the highest number of buildings vectorized with an accuracy of over 80%.