*1.1. CNN Deep Learning for Cadastral Mapping*

CNNs are one of the most popular and successful deep networks for image interpretation tasks. They are proven to work efficiently to identify various objects in remote sensing imagery [12–16]. Comprehensive overviews contextualizing the evolution of deep learning and CNNs in geoscience and remote sensing are provided by Bergen et al. and Zhu et al. [17,18]. In essence, CNNs are neural networks that incorporate the convolution and pooling operation as a layer. CNNs have been characterized by five concepts [19]:


In deep learning, there are two approaches to train a CNN: From scratch or via transfer learning [20]. When trained from scratch, all features are learned from data to be provided, which demands large amounts of data and comes with a higher risk of overfitting. An over-fitted network can make accurate predictions for a certain dataset, but fails to generalize its learning capacity for another dataset. With transfer learning, part of the features are learned from a different, typically large dataset. These low-level features are more general and abstract. The network has proven excellence for a specific application. Its core architecture is kept and applied to a new application. Only the last convolution block is trained on specific data of the new application, resulting in specialized high-level features. Transfer learning requires learning fewer features, and thus fewer data. In our study, we investigate transfer learning as an existing CNN for cadastral mapping.
