Abstract
Maps are an important medium that enable people to comprehensively understand the configuration of cultural activities and natural elements over different times and places. Although massive maps are available in the digital era, how to effectively and accurately access the required map remains a challenge today. Previous works partially related to map-type classification mainly focused on map comparison and map matching at the local scale. The features derived from local map areas might be insufficient to characterize map content. To facilitate establishing an automatic approach for accessing the needed map, this paper reports our investigation into using deep learning techniques to recognize seven types of map, including topographic map, terrain map, physical map, urban scene map, the National Map, 3D map, nighttime map, orthophoto map, and land cover classification map. Experimental results show that the state-of-the-art deep convolutional neural networks can support automatic map-type classification. Additionally, the classification accuracy varies according to different map-types. We hope our work can contribute to the implementation of deep learning techniques in cartographical community and advance the progress of Geographical Artificial Intelligence (GeoAI).
Introduction
In the digital age, we have access to an overwhelming volume of geospatial data, yet finding the specific type of map required remains difficult. Most existing map search and retrieval systems rely heavily on metadata—titles, tags, and legends. However, a vast majority of maps found online are "stripped" of these elements, leaving only the map frame itself. This research addresses this challenge by applying "Geographical Artificial Intelligence" (GeoAI) to perform global, image-level classification, moving beyond previous local-scale matching methods that often fail to capture the broader context of map content.
The deepMap Dataset
The researchers introduced a new benchmark called deepMap, curated from various online sources including ArcGIS, Google Maps, and the USGS National Map. The study utilized a dataset of 1,812 images (256×256 pixels) categorized into seven primary classes.
These categories include: (1) Topographic/Terrain/Physical maps, (2) Urban scene maps, (3) The National Map (USGS style), (4) 3D maps, (5) Nighttime imagery maps, (6) Orthophoto maps, and (7) Land cover classification maps. Standard preprocessing, including rescaling and data augmentation through rotation, was used to improve model robustness.
Methodology
The researchers treated map-type classification as a standard image recognition task, leveraging the power of Deep Convolutional Neural Networks (DCNNs) to extract high-level representations. They benchmarked five established architectures:
- AlexNet: A pioneering shallow deep network.
- VGG Net-19: A deeper model emphasizing network depth.
- Inception V4: Designed for computational efficiency and high accuracy.
- ResNet V2-152: A very deep network using residual blocks.
- ResNet-Inception V2: A hybrid model combining the strengths of both ResNet and Inception.
Experimental Results
The experimental results were highly promising, with classification accuracies ranging from 70% to 99% depending on the model and the volume of training data.
Key insights from the experiments include:
- Top Performer: The ResNet-Inception V2 model achieved the highest accuracy, reaching 95%–99% when trained on 80% of the dataset. Its hybrid architecture effectively captured the nuanced visual textures of different map types.
- Data Sensitivity: Moving from a 60/40 to an 80/20 training-test split consistently boosted accuracy across all models, sometimes by as much as 7%.
- Visual Complexity: Maps are often harder to classify than natural photographs because different types share a similar visual "vocabulary" (e.g., blue for water). The global context provided by DCNNs is essential for distinguishing these subtle differences.
Conclusion
While DCNNs show great potential for automated map classification, challenges remain regarding inter-class similarity, such as distinguishing "The National Map" from general topographic styles. Future work will investigate transfer learning and larger-scale datasets to further refine these models, advancing the progress of GeoAI in the cartographical community.
Summary updated