Change Detection and Semantic Segmentation of Historical Maps to Detect Indicators of Soil Contamination

Abstract

Historical maps are commonly used as a source of information to find indicators of soil contamination. The Netherlands has an extensive collection of historical maps, which makes it possible to analyse maps of any location from up to 200 years in the past. Because the analysis of the maps is time-consuming, an automated solution is sought. This thesis makes the first step in this regard by semantically segmenting 90 years of historical maps and detecting changes in the landscape per decade. Automatic change detection with historical maps is a novelty. The methods for this are based on work on remote sensing data. To perform both segmentation and change detection, one general convolutional neural network architecture is proposed, of which several variations are tested. The variations serve to distinguish and compare three main influences on the performance: the first is the influence of multi-temporal inputs on semantic segmentation. Adding the map from a decade after the segmented map to the input did not result in significantly higher performances overall; it is primarily useful in situations where a feature is drawn ambiguously while it is drawn clearly in the decade after it. The second influence is the chosen method to combine multiple maps. The two methods that were tried, namely concatenation and Convolutional Gated Recurrent Units, did not differ in performance for any task. Concatenation is therefore concluded to be the superior method, as it is faster and simpler. The final influence is the choice of performing the two tasks simultaneously or individually. Performing both tasks improved the change detection score while decreasing the segmentation score. Overall, the models learned to generalise features on maps from different decades and locations, with a semantic segmentation F1 score of 88% and a change detection F1 score of 72%. These scores are significantly higher than those of the baseline models based on random forests, which shows the added value of using neural networks for tasks that benefit from context information.