GisGCN: A Visual Graph-Based Framework to Match Geographical Areas through Time

Margarita Khokhlova,Nathalie Abadie,Liming Chen,Valérie Gouet-Brunet

doi:10.3390/ijgi11020097

Abstract

Historical visual sources are particularly useful for reconstructing the successive states of the territory in the past and for analysing its evolution. However, finding visual sources covering a given area within a large mass of archives can be very difficult if they are poorly documented. In the case of aerial photographs, most of the time, this task is carried out by solely relying on the visual content of the images. Convolutional Neural Networks are capable to capture the visual cues of the images and match them to each other given a sufficient amount of training data. However, over time and across seasons, the natural and man-made landscapes may evolve, making historical image-based retrieval a challenging task. We want to approach this cross-time aerial indexing and retrieval problem from a different novel point of view: by using geometrical and topological properties of geographic entities of the researched zone encoded as graph representations which are more robust to appearance changes than the pure image-based ones. Geographic entities in the vertical aerial images are thought of as nodes in a graph, linked to each other by edges representing their spatial relationships. To build such graphs, we propose to use instances from topographic vector databases and state-of-the-art spatial analysis methods. We demonstrate how these geospatial graphs can be successfully matched across time by means of the learned graph embedding.

Highlights

Historical visual sources, such as maps, engravings, drawings, or photographs, are the only visual representations of the past geographical realm still accessible
While our approach targets historical aerial image matching with georeferenced photographs, we propose a fundamentally different and novel way to encode the information captured by photographs: we propose to leverage the geometrical and topological properties of geographic entities represented by the photographs instead of pure visual cues
We follow the similar idea to use the descriptive power of graph representation along with a Siamese-based Graph Convolutional Network (GCN); the graph creation process differs from region adjacency graph (RAG) approach [69], and the architecture we propose is conceived for our type of data and corresponding features

Summary

Introduction

Historical visual sources, such as maps, engravings, drawings, or photographs, are the only visual representations of the past geographical realm still accessible As such, they are extremely important sources of information for analyzing the past territory and its evolution. A first challenge faced by these works is the georeferencing of the historical visual sources This operation aims at determining the geographic position of a given visual source. Whether performed manually [2] or automatically [3], this task requires providing enough ground control points to the system that estimates the georeferencing model This implies identifying beforehand the area covered by each historical visual source, which is, in the case of old aerial photographs, sometimes very poorly documented, can be extremely difficult. Graph-based representation of places and landscapes can reveal important insights in scenarios such as scene geolocalization or geographic information retrieval

Methods

Findings

Discussion

Conclusion