Computational Cartographic Recognition: Identifying Maps, Geographic Regions, and Projections from Images Using Machine Learning

Map reading is a challenging task for computer programs. This article explores how artificial intelligence and machine learning methods can be used to understand maps, an area we broadly refer to as computational cartographic recognition. Specifically, we use machine learning methods to (1) identify whether an image is a map, (2) recognize the geographic region on the map, and (3) recognize the projection used on the map. Four machine learning models—support vector machine, multilayer perceptrons, convolutional neural networks (CNNs) developed from scratch using our own architecture (CNNS), and pretrained CNN models through transfer learning (CNNT)—are applied in these tasks. We use 2,200 online map images, 500 nonmap images, and 1,050 synthetic map images to train and evaluate the models. Results show that the CNNT models achieve the highest performance among all models, with an accuracy rate above 90 percent for the tasks. The CNNS models come in second. We also conduct a round of stress tests using 3,600 additional synthetic maps where the shape and layout are systematically distorted and test if the models can still identify the maps and recognize the region and projection on the maps. The results of the stress tests show that the models can reliably recognize some of the modified maps even when exhibiting performance inferior to even random models for other maps. This unpredictable nature of the methods when applied to maps that are not represented in the training data suggests both promises and limitations of the current machine learning approaches to cartographic recognition.