Convolutional Neural Networks for Land Use and Land Cover Multi-class Maps from Historical Aerial Photographs

Historical maps that describe past land use and land cover (LULC) forms can be a precious source of information in many scientific fields studying long-term spatial and temporal changes in the landscape. Such repositories were created manually in small areas in the past, which was a time-consuming and labor-intensive task. Recently, there has been a growing tendency to use machine learning models for this purpose, along with deep learning methods. However, having a massive amount of labeled data is necessary for these methods to train the networks. Training data are often manually labeled, posing a significant challenge and limiting the automation of these methods. This article presents a method that uses topographic databases to extract complex multi-class maps representing LULC from historical aerial photographs, eliminating the time-consuming data labeling step. The method uses transfer learning with a pretrained model on 2020 and 2014 data and attempts to reconstruct LULC types with the same convolutional neural network (CNN) network on archived images from 2006. The experiment covered 488 km 2 and included seven LULC classes. The method was tested using different CNN architectures (U-Net, Pyramid Scene Parsing Network [PSPNet], and LinkNet) with backbones (ResNeXt+SE, EfficientNet, and Inception). The PSPNet‐EfficientNet‐b7 network model achieved the best results, with 90% overall accuracy for predicting LULC classes based on the 2006 archived aerial images.