Exploring geo-transferability of deep neural network by developing comprehensive metrics

Geo-transferability is an important but challenging dimension of the generalizability of GeoAI models because of the inherent geographic heterogeneity. Geographic similarity serves as a potential basis for enhancing the geo-transferability of GeoAI models. This study presents a novel approach to explore the geo-transferability of deep neural network (DNN) models for geographic tasks by developing comprehensive metrics. The geo-related DNN model is trained in the geographic source domain and transferred to the target domain across diverse geographic regions. Two comprehensive metrics, named cross-domain stability and cross-domain adaptability, are proposed to integrate multiple evaluation metrics, i.e. accuracy, precision, recall, and intersection over union to measure the geo-transferability of DNNs. Finally, the relationship between geographic similarity and geo-transferability is examined. The presented approach was experimentally analyzed in the Guangdong-Hong Kong-Macao Greater Bay Area by taking the urban village recognition as an example. The results show that the proposed metrics can be used to comprehensively measure the geo-transferability of Geo-DNN models. A positive correlation is revealed between geographic similarity and geo-transferability in the urban village recognition task. This study provides valuable insights into the development of GeoAI and transfer learning of DNN models and supports the further application of geospatial artificial intelligence.