TransMI: a transfer-learning method for generalized map information evaluation

Information Theory of Cartography plays an important role in guiding map design, generalization, and evaluation, and measurement of map information is the most basic topic of this theory. However, there are many problems in current measurement methods, and there is a long way to go to form a theoretically rigorous algorithm that can effectively depict spatial information and comprehensively consider the feeling of map readers. Luckily, we can now propose an evaluation metric that exhibits a certain correlation with map information based on deep learning to benefit actual cartography, and we term it as generalized map information evaluation to demonstrate differentiation. Specifically, this paper first constructs a subjective data set to support the supervised learning paradigm. Also, considering the difficulty of large-scale subjective data set collection, this paper proposes a Transfer-learning method for generalized Map Information evaluation (TransMI). Technically, a Siamese Network is pre-trained to explicitly acquire prior knowledge about the reasons for changes to mapped information. On this basis, one branch of the network is extracted and fine-tuned on the subjective data set to achieve the goal of predicting the quality of generalized map information. The results and the analysis of ablation studies prove the feasibility of our method.